Phillip Compeau, Pavel Pevzner - Bioinformatics Algorithms - An Active Learning Approach. Vol2 (2015)

Welcome!
Thank you for joining us! As you explore this book, you will find a number of active
learning components that help you learn the material at your own pace.
1. CODE CHALLENGES ask you to implement the algorithms that you will en-
counter (in any programming language you like). These code challenges are
hosted in the “Bioinformatics Textbook Track” location on Rosalind (http://
rosalind.info), a website that will automatically test your implementations.
2. CHARGING STATIONS provide additional insights on implementing the algo-

rithms you encounter. However, we suggest trying to solve a Code Challenge
before you visit a Charging Station.
3. EXERCISE BREAKS offer “just in time” assessments testing your understanding

of a topic before moving to the next one.
4. STOP and Think questions invite you to slow down and contemplate the current
material before continuing to the next topic.
5. DETOURS provide extra content that didn’t quite fit in the main text. DETOUR
6. FINAL CHALLENGES ask you to apply what you have learned to real experi-
mental datasets.
This textbook powers our popular online courses on Coursera. We encourage you
to sign up for a session and learn this material while interacting with thousands of
other talented students from around the world. You can also find lecture videos and
PowerPoint slides at the textbook website, http://bioinformaticsalgorithms.org.
i
Bioinformatics Algorithms:
An Active Learning Approach
2nd Edition, Vol. II
Phillip Compeau & Pavel Pevzner
http://bioinformaticsalgorithms.org
© 2015
Copyright © 2015 by Phillip Compeau and Pavel Pevzner. All rights reserved.
This book or any portion thereof may not be reproduced or used in any manner whatso-
ever without the express written permission of the publisher except for the use of brief
quotations in a book review.
Printed in the United States of America
First Printing, 2015
ISBN: 978-0-9903746-2-6
Library of Congress Control Number: 2015945208
Active Learning Publishers, LLC

9768 Claiborne Square
La Jolla, CA 92037
To my family. — P. C.
To my parents. — P. P.
In Case You Missed Volume I. . .
C HAPTER 1 C HAPTER 2
vi
vii
Volume II Overview
C HAPTER 7 — p. 2 C HAPTER 8 — p. 68
viii
CHAPTER 9 — p. 120 CHAPTER 10 — p. 178
CHAPTER 11 — p. 234
ix
Contents
List of Code Challenges xvii
About the Textbook xix

Meet the Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix
Meet the Development Team . . . . . . . . . . . . . . . . . . . . . . . . . . . . xx
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi
7 Which Animal Gave Us SARS? 2

The Fastest Outbreak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Trouble at the Metropole Hotel . . . . . . . . . . . . . . . . . . . . . . . . 3
The evolution of SARS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Transforming Distance Matrices into Evolutionary Trees . . . . . . . . . . . . . 5
Constructing a distance matrix from coronavirus genomes . . . . . . . . 5
Evolutionary trees as graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Distance-based phylogeny construction . . . . . . . . . . . . . . . . . . . 10
Toward An Algorithm for Distance-Based Phylogeny Construction . . . . . . 12
A quest for neighboring leaves . . . . . . . . . . . . . . . . . . . . . . . . 12
Computing limb lengths . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Additive Phylogeny . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Trimming the tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Attaching a limb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
An algorithm for distance-based phylogeny construction . . . . . . . . . 19
Constructing an evolutionary tree of coronaviruses . . . . . . . . . . . . 20
Using Least Squares to Construct Approximate Distance-Based Phylogenies . 22
Ultrametric Evolutionary Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
The Neighbor-Joining Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Transforming a distance matrix into a neighbor-joining matrix . . . . . . 27
x
Analyzing coronaviruses with the neighbor-joining algorithm . . . . . . 31
Limitations of distance-based approaches to evolutionary tree construction 33
Character-Based Tree Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . 33
Character tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
From anatomical to genetic characters . . . . . . . . . . . . . . . . . . . . 34
How many times has evolution invented insect wings? . . . . . . . . . . 35
The Small Parsimony Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
The Large Parsimony Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Epilogue: Evolutionary Trees Fight Crime . . . . . . . . . . . . . . . . . . . . . 48
Detours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
When did HIV jump from primates to humans? . . . . . . . . . . . . . . 51
Searching for a tree fitting a distance matrix . . . . . . . . . . . . . . . . . 51
The four point condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Did bats give us SARS? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Why does the neighbor-joining algorithm find neighboring leaves? . . . 56
Computing limb lengths in the neighbor-joining algorithm . . . . . . . . 61
Giant panda: bear or raccoon? . . . . . . . . . . . . . . . . . . . . . . . . . 62
Where did humans come from? . . . . . . . . . . . . . . . . . . . . . . . . 62
Bibliography Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
8 How Did Yeast Become a Wine Maker? 68

An Evolutionary History of Wine Making . . . . . . . . . . . . . . . . . . . . . 69
How long have we been addicted to alcohol? . . . . . . . . . . . . . . . . 69
The diauxic shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Identifying Genes Responsible for the Diauxic Shift . . . . . . . . . . . . . . . 70
Two evolutionary hypotheses with different fates . . . . . . . . . . . . . 70
Which yeast genes drive the diauxic shift? . . . . . . . . . . . . . . . . . . 71
Introduction to Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Gene expression analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Clustering yeast genes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
The Good Clustering Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Clustering as an Optimization Problem . . . . . . . . . . . . . . . . . . . . . . 78
Farthest First Traversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
k-Means Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Squared error distortion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
k-means clustering and the center of gravity . . . . . . . . . . . . . . . . . 83
The Lloyd Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
xi
From centers to clusters and back again . . . . . . . . . . . . . . . . . . . 85
Initializing the Lloyd algorithm . . . . . . . . . . . . . . . . . . . . . . . . 87
k-means++ Initializer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Clustering Genes Implicated in the Diauxic Shift . . . . . . . . . . . . . . . . . 89
Limitations of k-Means Clustering . . . . . . . . . . . . . . . . . . . . . . . . . 90
From Coin Flipping to k-Means Clustering . . . . . . . . . . . . . . . . . . . . 92
Flipping coins with unknown biases . . . . . . . . . . . . . . . . . . . . . 92
Where is the computational problem? . . . . . . . . . . . . . . . . . . . . 95
From coin flipping to the Lloyd algorithm . . . . . . . . . . . . . . . . . . 95
Return to clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Making Soft Decisions in Coin Flipping . . . . . . . . . . . . . . . . . . . . . . 97
Expectation maximization: the E-step . . . . . . . . . . . . . . . . . . . . 97
Expectation maximization: the M-step . . . . . . . . . . . . . . . . . . . . 99
The expectation maximization algorithm . . . . . . . . . . . . . . . . . . 100
Soft k-Means Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Applying expectation maximization to clustering . . . . . . . . . . . . . 100
Centers to soft clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Soft clusters to centers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Hierarchical Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Introduction to distance-based clustering . . . . . . . . . . . . . . . . . . 103
Inferring clusters from a tree . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Analyzing the diauxic shift with hierarchical clustering . . . . . . . . . . 108
Epilogue: Clustering Tumor Samples . . . . . . . . . . . . . . . . . . . . . . . . 109
Detours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Whole genome duplication or a series of duplications? . . . . . . . . . . 111
Measuring gene expression . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Microarrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Proof of the Center of Gravity Theorem . . . . . . . . . . . . . . . . . . . 113
Transforming an expression matrix into a distance/similarity matrix . . 114
Clustering and corrupted cliques . . . . . . . . . . . . . . . . . . . . . . . 115
9 How Do We Locate Disease-Causing Mutations? 120

What Causes Ohdo Syndrome? . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Introduction to Multiple Pattern Matching . . . . . . . . . . . . . . . . . . . . . 122
Herding Patterns into a Trie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Constructing a trie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
xii
Applying the trie to multiple pattern matching . . . . . . . . . . . . . . . 125
Preprocessing the Genome Instead . . . . . . . . . . . . . . . . . . . . . . . . . 127
Introduction to suffix tries . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Using suffix tries for pattern matching . . . . . . . . . . . . . . . . . . . . 127
Suffix Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Suffix Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Constructing a suffix array . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Pattern matching with the suffix array . . . . . . . . . . . . . . . . . . . . 134
The Burrows-Wheeler Transform . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Genome compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Constructing the Burrows-Wheeler transform . . . . . . . . . . . . . . . . 136
From repeats to runs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
Inverting the Burrows-Wheeler Transform . . . . . . . . . . . . . . . . . . . . . 139
A first attempt at inverting the Burrows-Wheeler transform . . . . . . . 139
The First-Last Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Using the First-Last property to invert the Burrows-Wheeler transform . 144
Pattern Matching with the Burrows-Wheeler Transform . . . . . . . . . . . . . 147
A first attempt at Burrows-Wheeler pattern matching . . . . . . . . . . . 147
Moving backward through a pattern . . . . . . . . . . . . . . . . . . . . . 148
The Last-to-First mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Speeding Up Burrows-Wheeler Pattern Matching . . . . . . . . . . . . . . . . 153
Substituting the Last-to-First mapping with count arrays . . . . . . . . . 153
Getting rid of the first column of the Burrows-Wheeler matrix . . . . . . 154
Where are the Matched Patterns? . . . . . . . . . . . . . . . . . . . . . . . . . . 156
Burrows and Wheeler Set Up Checkpoints . . . . . . . . . . . . . . . . . . . . . 157
Epilogue: Mismatch-Tolerant Read Mapping . . . . . . . . . . . . . . . . . . . 159
Reducing approximate pattern matching to exact pattern matching . . . 159
BLAST: Comparing a sequence against a database . . . . . . . . . . . . . 160
Approximate pattern matching with the Burrows-Wheeler transform . . 162
Charging Stations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Constructing a suffix tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Solving the Longest Shared Substring Problem . . . . . . . . . . . . . . . 167
Partial suffix array construction . . . . . . . . . . . . . . . . . . . . . . . . 169
Detours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
The reference human genome . . . . . . . . . . . . . . . . . . . . . . . . . 170
Rearrangements, insertions, and deletions in human genomes . . . . . . 170
The Aho-Corasick algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 170
xiii
From suffix trees to suffix arrays . . . . . . . . . . . . . . . . . . . . . . . 171
From suffix arrays to suffix trees . . . . . . . . . . . . . . . . . . . . . . . 173
Binary search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
10 Why Have Biologists Still Not Developed an HIV Vaccine? 178

Classifying the HIV Phenotype . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
How does HIV evade the human immune system? . . . . . . . . . . . . . 179
Limitations of sequence alignment . . . . . . . . . . . . . . . . . . . . . . 181
Gambling with Yakuza . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Two Coins up the Dealer’s Sleeve . . . . . . . . . . . . . . . . . . . . . . . . . . 184
Finding CG-Islands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Hidden Markov Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
From coin flipping to a Hidden Markov Model . . . . . . . . . . . . . . . 186
The HMM diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
Reformulating the Casino Problem . . . . . . . . . . . . . . . . . . . . . . 188
The Decoding Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
The Viterbi graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
The Viterbi algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
How fast is the Viterbi algorithm? . . . . . . . . . . . . . . . . . . . . . . 195
Finding the Most Likely Outcome of an HMM . . . . . . . . . . . . . . . . . . 196
Profile HMMs for Sequence Alignment . . . . . . . . . . . . . . . . . . . . . . . 198
How do HMMs relate to sequence alignment? . . . . . . . . . . . . . . . 198
Building a profile HMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Transition and emission probabilities of a profile HMM . . . . . . . . . . 203
Classifying proteins with profile HMMs . . . . . . . . . . . . . . . . . . . . . . 207
Aligning a protein against a profile HMM . . . . . . . . . . . . . . . . . . 207
The return of pseudocounts . . . . . . . . . . . . . . . . . . . . . . . . . . 208
The troublesome silent states . . . . . . . . . . . . . . . . . . . . . . . . . 209
Are profile HMMs really all that useful? . . . . . . . . . . . . . . . . . . . 216
Learning the Parameters of an HMM . . . . . . . . . . . . . . . . . . . . . . . . 217
Estimating HMM parameters when the hidden path is known . . . . . . 217
Viterbi learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Soft Decisions in Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . 221
The Soft Decoding Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 221
The forward-backward algorithm . . . . . . . . . . . . . . . . . . . . . . . 222
Baum-Welch Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
xiv
The Many Faces of HMMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
Epilogue: Nature is a Tinkerer and not an Inventor . . . . . . . . . . . . . . . 227
Detours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
The Red Queen Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Glycosylation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
DNA methylation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Conditional probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
11 Was T. rex Just a Big Chicken? 234

Paleontology Meets Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Which Proteins Are Present in This Sample? . . . . . . . . . . . . . . . . . . . . 236
Decoding an Ideal Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
From Ideal to Real Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Peptide Sequencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
Scoring peptides against spectra . . . . . . . . . . . . . . . . . . . . . . . 244
Where are the suffix peptides? . . . . . . . . . . . . . . . . . . . . . . . . . 246
Peptide sequencing algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 248
Peptide Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
The Peptide Identification Problem . . . . . . . . . . . . . . . . . . . . . . 249
Identifying peptides in the unknown T. rex proteome . . . . . . . . . . . 250
Searching for peptide-spectrum matches . . . . . . . . . . . . . . . . . . . 251
Peptide Identification and the Infinite Monkey Theorem . . . . . . . . . . . . . 252
False discovery rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
The monkey and the typewriter . . . . . . . . . . . . . . . . . . . . . . . . 254
Statistical significance of a peptide-spectrum match . . . . . . . . . . . . 255
Spectral Dictionaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
T. rex Peptides: Contaminants or Treasure Trove of Ancient Proteins? . . . . . 261
The hemoglobin riddle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
The dinosaur DNA controversy . . . . . . . . . . . . . . . . . . . . . . . . 264
Epilogue: From Unmodified to Modified Peptides . . . . . . . . . . . . . . . . 264
Post-translational modifications . . . . . . . . . . . . . . . . . . . . . . . . 264
Searching for modifications as an alignment problem . . . . . . . . . . . 265
Building a Manhattan grid for spectral alignment . . . . . . . . . . . . . 267
Spectral alignment algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 271
Detours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Gene prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
xv
Finding all paths in a graph . . . . . . . . . . . . . . . . . . . . . . . . . . 274
The Anti-Symmetric Path Problem . . . . . . . . . . . . . . . . . . . . . . 275
Transforming spectra into spectral vectors . . . . . . . . . . . . . . . . . . 276
The infinite monkey theorem . . . . . . . . . . . . . . . . . . . . . . . . . 278
The probabilistic space of peptides in a spectral dictionary . . . . . . . . 278
Are terrestrial dinosaurs really the ancestors of birds? . . . . . . . . . . . 279
Solving the Most Likely Peptide Vector Problem . . . . . . . . . . . . . . 280
Selecting parameters for transforming spectra into spectral vectors . . . 281
Bibliography 285
Image Courtesies 291
xvi
List of Code Challenges
Chapter 7 2
(7A) Compute Distances Between Leaves . . . . . . . . . . . . . . . . . . . . . 11
(7B) Compute Limb Lengths in a Tree . . . . . . . . . . . . . . . . . . . . . . . 17
(7C) Implement A DDITIVE P HYLOGENY . . . . . . . . . . . . . . . . . . . . . . 20
(7D) Implement UPGMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
(7E) Implement N EIGHBOR J OINING . . . . . . . . . . . . . . . . . . . . . . . 30
(7F) Implement S MALL PARSIMONY . . . . . . . . . . . . . . . . . . . . . . . . 40
(7G) Adapt S MALL PARSIMONY to Unrooted Trees . . . . . . . . . . . . . . . . 42
(7H) Find the Nearest Neighbors of a Tree . . . . . . . . . . . . . . . . . . . . . 45
(7I) Implement N EAREST N EIGHBOR I NTERCHANGE . . . . . . . . . . . . . . 47
Chapter 8 68
(8A) Implement FARTHEST F IRST T RAVERSAL . . . . . . . . . . . . . . . . . . 80
(8B) Compute the Squared Error Distortion . . . . . . . . . . . . . . . . . . . . 82
(8C) Implement the Lloyd Algorithm for k-Means Clustering . . . . . . . . . . 85
(8D) Implement the Soft k-Means Clustering Algorithm. . . . . . . . . . . . . 103
(8E) Implement H IERARCHICAL C LUSTERING . . . . . . . . . . . . . . . . . . 106
Chapter 9 120
(9A) Construct a Trie from a Collection of Patterns . . . . . . . . . . . . . . . . 124
(9B) Implement T RIE M ATCHING . . . . . . . . . . . . . . . . . . . . . . . . . 126
(9C) Construct the Suffix Tree of a String . . . . . . . . . . . . . . . . . . . . . 132
(9D) Find the Longest Repeat in a String . . . . . . . . . . . . . . . . . . . . . . 132
(9E) Find the Longest Substring Shared by Two Strings . . . . . . . . . . . . . 133
(9F) Find the Shortest Non-Shared Substring of Two Strings . . . . . . . . . . 133
(9G) Construct the Suffix Array of a String . . . . . . . . . . . . . . . . . . . . 133
(9H) Implement PATTERN M ATCHING W ITH S UFFIX A RRAY . . . . . . . . . . . 135
xvii
(9I) Construct the Burrows-Wheeler Transform of a String . . . . . . . . . . . 138
(9J) Reconstruct a String from its Burrows-Wheeler Transform . . . . . . . . 147
(9K) Generate the Last-to-First Mapping of a String . . . . . . . . . . . . . . . 151
(9L) Implement BWM ATCHING . . . . . . . . . . . . . . . . . . . . . . . . . . 151
(9M) Implement B ETTER BWM ATCHING . . . . . . . . . . . . . . . . . . . . . 156
(9N) Find All Occurrences of a Collection of Patterns in a String . . . . . . . . 158
(9O) Find All Approximate Occurrences of a Collection of Patterns in a String 162
(9P) Implement T REE C OLORING . . . . . . . . . . . . . . . . . . . . . . . . . 169
(9Q) Construct the Partial Suffix Array of a String . . . . . . . . . . . . . . . . 169
(9R) Construct a Suffix Tree from a Suffix Array . . . . . . . . . . . . . . . . . 172
Chapter 10 178
(10A) Compute the Probability of a Hidden Path . . . . . . . . . . . . . . . . . 190
(10B) Compute the Probability of an Outcome Given a Hidden Path . . . . . 191
(10C) Implement the Viterbi Algorithm . . . . . . . . . . . . . . . . . . . . . . 195
(10D) Compute the Probability of a String Emitted by an HMM . . . . . . . . 197
(10E) Construct a Profile HMM . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
(10F) Construct a Profile HMM with Pseudocounts . . . . . . . . . . . . . . . 209
(10G) Perform a Multiple Sequence Alignment with a Profile HMM . . . . . . 212
(10H) Estimate the Parameters of an HMM . . . . . . . . . . . . . . . . . . . . 219
(10I) Implement Viterbi Learning . . . . . . . . . . . . . . . . . . . . . . . . . 220
(10J) Solve the Soft Decoding Problem . . . . . . . . . . . . . . . . . . . . . . 223
(10K) Implement Baum-Welch Learning . . . . . . . . . . . . . . . . . . . . . . 226
Chapter 11 234
(11A) Construct the Graph of a Spectrum . . . . . . . . . . . . . . . . . . . . . 239
(11B) Implement D ECODING I DEAL S PECTRUM . . . . . . . . . . . . . . . . . 240
(11C) Convert a Peptide into a Peptide Vector . . . . . . . . . . . . . . . . . . 245
(11D) Convert a Peptide Vector into a Peptide . . . . . . . . . . . . . . . . . . 245
(11E) Sequence a Peptide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
(11F) Find a Highest-Scoring Peptide in a Proteome against a Spectrum . . . 250
(11G) Implement PSMS EARCH . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
(11H) Compute the Size of a Spectral Dictionary . . . . . . . . . . . . . . . . . 258
(11I) Compute the Probability of a Spectral Dictionary . . . . . . . . . . . . . 260
(11J) Find a Highest-Scoring Modified Peptide against a Spectrum . . . . . . 273
xviii
About the Textbook
Meet the Authors
P HILLIP C OMPEAU is an Assistant Teaching Professor in the Compu-

tational Biology Department at Carnegie Mellon University. He is a
former postdoctoral researcher in the Department of Computer Sci-
ence & Engineering at the University of California, San Diego, where
he received a Ph. D. in mathematics. He is passionate about the future
of both offline and online education, having cofounded Rosalind with
Nikolay Vyahhi in 2012. A retired tennis player, he dreams of one day
going pro in golf.
PAVEL P EVZNER is Ronald R. Taylor Professor of Computer Science

at the University of California, San Diego. He holds a Ph. D. from
Moscow Institute of Physics and Technology, Russia and an Hon-
orary Degree from Simon Fraser University. He is a Howard Hughes
Medical Institute Professor (2006), an Association for Computing Ma-
chinery Fellow (2010), and an International Society for Computational
Biology Fellow (2012). He has authored the textbooks Computational
Molecular Biology: An Algorithmic Approach (2000) and An Introduction
to Bioinformatics Algorithms (2004) (jointly with Neil Jones).
xix
Meet the Development Team
V U N GO is a Ph. D student in Bioinformatics and Systems Biology at

UCSD. He holds a B. S in Biotechnology from Rutgers University. His
research interests include epigenetics, gene regulatory networks, and
machine learning algorithms. Outside of research, Vu enjoys movies,
music and badminton.
M AX S HEN is a student in the Computational and Systems Biology

Ph. D. program at the Massachusetts Institute of Technology, having re-
ceived a B. S. from UCSD. He is excited about the ways algorithms and
machine learning are applied in bioinformatics. Outside of research,
Max enjoys hip hop dance and playing video games competitively.
J EFFREY Y UAN is a graduate student in the Bioinformatics and Sys-

tems Biology Program at UCSD. He holds a Sc. B. in Computational
Biology from Brown University. His research interests include genome
assembly with long reads, the 3-dimensional organization of genome
structure, and genomic variation. Outside of research, Jeffrey enjoys
reading, board games, and volleyball.
xx
Acknowledgments
This textbook was greatly improved by the efforts of a large number of individuals, to
whom we owe a debt of gratitude.
The development team (Vu Ngo, Max Shen, and Jeffrey Yuan), as well as Ksenia
Krasheninnikova, implemented coding challenges and exercises, rendered figures,
helped typeset the text, and offered insightful feedback on the manuscript.
Glenn Tesler provided thorough chapter reviews and caught many errors in our
manuscript.
Sangtae Kim, Seungjin Na, Mihai Pop, and Héctor Corrada Bravo provided many
thoughtful comments. Sangtae and Seungjin also generated some images used in the
book.
Randall Christopher brought to life our ideas for illustrations in addition to the
textbook cover.
Nikolay Vyahhi led a team composed of Andrey Balandin, Artem Suschev, Aleksey
Kladov, and Kirill Shikhanov, who worked hard to support an online, interactive version
of this textbook used in our online course on Coursera.
Laurence Bernstein and Kai Zhang worked to implement many of the problems in
Chapter 9.
Our students on Coursera, especially Mark Mammel and Erika Ramírez, found
hundreds of typos in our preliminary manuscript.
Howard Hughes Medical Institute, the Russian Ministry of Education and Science,
and the National Institutes of Health generously gave their support for the development
of the online course based on this textbook. The Bioinformatics and Systems Biology
Program and the Computer Science & Engineering Department at the University of
California, San Diego provided additional support.
Finally, our families gracefully endured the many long days and nights that we
spent poring over manuscripts, and they helped us preserve our sanity along the way.
P. C. and P. P.
San Diego
July 2015
xxi
W H I C H A N I M A L G AV E U S S A R S ?
The Fastest Outbreak
Trouble at the Metropole Hotel
On February 21, 2003, a Chinese doctor named Liu Jianlun flew to Hong Kong to attend
a wedding and checked into Room 911 of the Metropole Hotel. The next day, he became
too ill to attend the wedding and was admitted to a hospital. Two weeks later, Dr.
Jianlun was dead.
On his deathbed, Jianlun told doctors that he had recently treated sick patients in
Guangdong Province, China where a deadly, highly contagious respiratory illness had
infected hundreds of people. The Chinese government had made brief mention of this
incident to the World Health Organization but had concluded that the likely culprit was
a common bacterial infection.
By the time anyone realized the severity of the disease, it was already too late to stop
the outbreak. On February 23, a man who had stayed across the hall from Dr. Jianlun at
the Metropole traveled to Hanoi and died after infecting 80 people. On February 26, a
woman checked out of the Metropole, traveled back to Toronto, and died after initiating
an outbreak there. On March 1, a third guest was admitted to a hospital in Singapore,
where sixteen additional cases of the illness arose within two weeks.
Consider that it took four years for the Black Death, which killed over a third of all
Europeans in the 14th Century, to travel from Constantinople to Kiev. Or that HIV took
two decades to circle the globe. In contrast, this mysterious new disease had crossed
the Pacific Ocean within a week of entering Hong Kong.
As health officials braced for the impact of the fastest-traveling pandemic in human
history, panic set in. Businesses were closed, sick passengers were removed from
airplanes, and Chinese officials threatened to execute infected patients who violated
quarantine.
International travel may have helped the disease spread rapidly, but international
collaboration would eventually contain it. In a matter of a few weeks, biologists identi-
fied a virus that had caused the epidemic and sequenced its genome. In the process,
the mysterious new disease earned a name: Severe Acute Respiratory Syndrome, or
SARS.
The evolution of SARS
The virus causing SARS belongs to a family of viruses called coronaviruses, which are
named after the Latin corona (meaning “crown”) because the virus particle resembles
the sun’s corona (Figure 7.1). Coronaviruses infect the respiratory tracts of mammals
3
CHAPTER 7
and birds but typically cause only minor problems, like the common cold. Before SARS,
no one believed that a coronavirus could wreak such havoc.
FIGURE 7.1 (Left) Coronavirus particles. (Right) A solar eclipse with the sun’s corona
visible.
Coronaviruses, influenza viruses, and HIV are all RNA viruses, meaning that they
possess RNA instead of DNA. RNA replication has a higher error rate than DNA
replication, and so RNA viruses are capable of mutating more quickly into divergent
strains. The rapid mutation of RNA viruses explains why the flu shot changes from
year to year and why there are many different subtypes of HIV.
SARS researchers initially hypothesized that, like HIV and influenza, the SARS
coronavirus (abbreviated as SARS-CoV) had jumped from animals to humans. They
first named birds as the likely suspect because of the similarities between SARS and
“bird flu”, a form of influenza originating in chickens that is difficult to transmit to
humans but is even deadlier than SARS, killing over half of the people it infects. Yet
when researchers sequenced the 29,751 nucleotide-long SARS-CoV genome in April
2003, it became evident that SARS did not come from birds because its genome did not
resemble avian coronaviruses.
By fall 2003, researchers had sequenced many SARS-CoV strains from patients in
various countries, but many questions still remained unanswered. How did SARS-CoV
cross the species barrier to humans? When and where did it happen? How did SARS
spread around the world, and who infected whom?
Each of these questions about SARS is ultimately related to the problem of con-
structing evolutionary trees (also known as phylogenies). For another example, by
constructing an evolutionary tree of primate viruses related to HIV (Figure 7.2), sci-
entists inferred that HIV was transmitted to humans on five separate occasions (see
PAGE 51 DETOUR: When Did HIV Jump From Primates to Humans?). But what algorithm
4
did they use to construct this phylogeny?
human HIV/M
human HIV/M
chimpanzee SIV
chimpanzee SIV
human HIV/N
human HIV/N
chimpanzee SIV
chimpanzee SIV
chimpanzee SIV
chimpanzee SIV
chimpanzee SIV
chimpanzee SIV
chimpanzee SIV
human HIV/O
human HIV/O
chimpanzee SIV
chimpanzee SIV
red-capped manabey SIV
drill SIV
vervet monkey SIV
tantalus monkey SIV
sooty mangabey SIV
human HIV/A
human HIV/B
sooty mangabey SIV
Sykes’s monkey SIV
greater spot-nosed monkey SIV
De Brazzas monkey SIV
FIGURE 7.2 HIV comprises five different viral families, denoted as A, B, M, N, and
O, with the M family responsible for 95% of all HIV infections. The five families are
different offshoots of the evolutionary tree for Simian Immunodeficiency Virus (SIV),
which infects primates. Stars indicate viruses transitioning from primates to humans.
The A and B families originated in sooty mangabey monkeys, whereas the M, N, and
O families originated in chimpanzees.
Transforming Distance Matrices into Evolutionary Trees
Constructing a distance matrix from coronavirus genomes
To determine how SARS jumped from animals to humans, scientists started sequencing
coronaviruses from various species in order to determine which one is the most similar
5
CHAPTER 7
to SARS-CoV. However, constructing a multiple alignment of entire viral genomes is

tricky because viral genes are often rearranged, inserted, and deleted. For this reason,
scientists focused on only one of the six genes in SARS-CoV. This gene encodes the
Spike protein, which identifies and binds to receptor sites on the host’s cell membrane.
In SARS-CoV, the Spike protein is 1,255 amino acids long and has rather weak
similarity with Spike proteins in other coronaviruses. However, even these subtle
similarities turned out to be sufficient for constructing a multiple alignment of Spike
proteins across various coronaviruses.
After constructing a multiple alignment of genes from n different species, biologists
often transform this alignment into an n ⇥ n distance matrix D. In many cases, Di,j
represents the number of differing symbols between the genes representing rows i
and j of the alignment (Figure 7.3). However, distance matrices can be constructed
using a variety of different distance functions in order to suit different applications.
For example, Di,j could represent the edit distance between genes from the i-th and
j-th species. Or, a distance matrix for n genomes could be constructed from the 2-break
distances between each pair of genomes.
Regardless of which distance function we use, in order to be a distance matrix, D
must satisfy three properties. It must be symmetric (for all i and j, Di,j = D j,i ), non-
negative (for all i and j, Di,j 0) and satisfy the triangle inequality (for all i, j, and k,
Di,j + D j,k  Di,k ).
EXERCISE BREAK: Prove that if Di,j is equal to the number of differing symbols
between rows i and j of a multiple alignment, then D is symmetric, non-negative,
and satisfies the triangle inequality.
SPECIES ALIGNMENT DISTANCE MATRIX

Chimp Human Seal Whale
Chimp ACGTAGGCCT 0 3 6 4
Human ATGTAAGACT 3 0 7 5
Seal TCGAGAGCAC 6 7 0 2
Whale TCGAAAGCAT 4 5 2 0
FIGURE 7.3 A multiple alignment of hypothetical DNA sequences from four species,
along with the distance matrix produced by counting the number of differing symbols
between each pair of rows in this multiple alignment.
6
By the end of 2003, bioinformaticians had sequenced many coronaviruses taken from a
variety of animals and SARS patients and then computed the associated distance matrix.
They needed to use this information in order to construct a coronavirus phylogeny and
understand the origin and spread of the SARS epidemic.
Evolutionary trees as graphs
You may have noticed that the HIV tree in Figure 7.2 has the structure of a graph.
Furthermore, Figure 7.4 (top) shows the representation of the phylogeny of all life as a
graph.
Graphs that are used to model phylogenies share two properties. They are connected
(i.e. it is possible to reach any node from any other node), and they contain no cycles.
For this reason, we will define a tree as a connected graph without cycles (see Figure 7.4
(bottom) for a few additional examples).
Take another look at Figure 7.4 (top). You will see that present-day species have
been assigned to the leaves of the tree, or nodes having degree 1 (in Chapter 3, we
defined the degree of a node as the number of edges connected to that node). Nodes
with degree larger than 1 are called internal nodes and represent unknown ancestor
species. Given a leaf j, there is only one node connected to j by an edge, which we call
the parent of j, denoted PARENT ( j). An edge connecting a leaf to its parent is called a
limb.
EXERCISE BREAK: Prove the following statements:
• Every tree with at least two nodes has at least two leaves.
• Every tree with n nodes has n 1 edges.
In a rooted tree, one node is designated as a special node called the root, and the edges
in the tree automatically inherit an implicit orientation away from the root, which is
placed at the top or left of the tree (Figure 7.5). This edge orientation models time: the
ancestor of all species in the tree is found at the root, and evolution proceeds from the
root outward through the tree. Trees without a designated root are called unrooted.
STOP and Think: Where would you place the root in the phylogeny in Fig-
ure 7.2?
7
CHAPTER 7
LIFE
bacteria
EUKARYOTES
archaebacteria
protoctists
PLANTS
ANIMALS
green algae fungi
mosses sponges
ferns cnidarians
owering owering worms

seed plants seed plants
lophophorates rotifers roundworms
VERTEBRATES ARTHROPODS
echinoderms
cartilaginous segmented mollusks chelicerates

worms
TETRAPODS
bon crustaceans insects
AMNIOTES
amphibians
mammals
turtles
snakes crocodiles
& lizards & birds
FIGURE 7.4 (Top) A connected acyclic graph that models an evolutionary tree of life
on Earth. Present-day species are shown as darker nodes (leaves). (Bottom) Trees come
in a variety of different shapes. In each of the three trees shown, leaves (i.e., nodes of
degree 1) have been drawn darker than internal nodes (i.e., nodes of larger degree).
8
We will analyze rooted trees when we attempt to infer the node corresponding to the
ancestor of all species in the tree; otherwise, we will analyze unrooted trees. Figure 7.6
shows an unrooted tree of HIV viruses produced from a different dataset than the one
used to create Figure 7.2. By proposing two additional subtypes of HIV, it illustrates
that the classification of HIV into five families shown in Figure 7.2 is not written in stone.
Ancestor
TIME
Present Day
FIGURE 7.5 A rooted tree, with the root (representing an ancestor of all species in the
tree) indicated in green at the top of the tree. The presence of the root implies an
orientation of edges in the tree away from the root.
SIV 1
HIV A SIV 2
SIV 7 HIV M
SIV 6 HIV N
HIV G HIV O
HIV B SIV 3
HIV F SIV 4
SIV 5
FIGURE 7.6 An unrooted tree of HIV and SIV viruses that suggests additional viral
families F and G in addition to the viral families A, B, M, N, and O shown in Figure 7.2.
9
CHAPTER 7
Distance-based phylogeny construction
We will first focus on deriving an unrooted tree from a distance matrix. The leaves
of this tree should correspond to the species represented by the matrix (with internal
nodes corresponding to unknown ancestral species). To reflect the evolutionary distance
between species in a tree, we assign each edge a non-negative length representing the
distance between the organisms that the edge connects, as shown in Figure 7.7.
Seal
0.5
Chimp
1 1.5
3
1 0
Whale
Human
1
Chimp Seal
1 2
3
2 0
Human Whale
FIGURE 7.7 Two unrooted trees fitting the distance matrix in Figure 7.3. Each of the five
maximal non-branching paths in the tree on the top is shown using a different color.
Replacing each maximal non-branching path in this tree with a single edge (of length
equal to the total length of edges) results in the simple tree shown on the bottom.
EXERCISE BREAK: Prove that there exists exactly one path connecting every
pair of nodes in a tree. Hint: what would happen if there were two different
paths connecting a pair of nodes? What would happen if there were no paths
connecting a pair of nodes?
In this chapter, we define the length of a path in a tree as the sum of the lengths of
its edges (rather than the number of edges on the path). As a result, the evolutionary
distance between two present-day species corresponding to leaves i and j in a tree T is
equal to the length of the unique path connecting i and j, denoted di,j ( T ).
10
Distances Between Leaves Problem:

Compute the distances between leaves in a weighted tree.
Input: A weighted tree with n leaves. 7A

Output: An n ⇥ n matrix (di,j ), where di,j is the length of the path between
leaves i and j.
The Distance Between Leaves Problem is straightforward to solve, but we would like to
solve the reverse problem, in which we must construct an unrooted tree that models a
given distance matrix. We say that a weighted unrooted tree T fits a distance matrix D
if di,j ( T ) = Di,j for every pair of leaves i and j.
Distance-Based Phylogeny Problem:

Reconstruct an evolutionary tree fitting a distance matrix.
Input: A distance matrix.

Output: A tree fitting this distance matrix.
STOP and Think: Does the Distance-Based Phylogeny Problem always have a
solution?
Not every distance matrix has a tree fitting it (see DETOUR: Searching for a Tree
Fitting a Distance Matrix). We therefore call a distance matrix additive if there exists PAGE 51
a tree that fits this matrix and non-additive otherwise. The term “additive” is used
because the lengths of all edges along the path between leaves i and j in a tree fitting
the matrix D add to Di,j .
Note that both trees in Figure 7.7 fit the distance matrix from Figure 7.3, so it would
be nice to have a notion of a “canonical” tree fitting a distance matrix. Extending
definitions introduced in Chapter 3 to undirected graphs, we say that a path in a tree is
non-branching if every node other than the beginning and ending node of the path has
degree equal to 2. A non-branching path is maximal if it is not a subpath of an even
longer non-branching path. If we substitute every maximal non-branching path by a
single edge whose length is equal to the length of the path, then the tree in Figure 7.7
(top) becomes the tree in Figure 7.7 (bottom). In general, after such a transformation,
there are no nodes of degree 2; a tree satisfying this property is called a simple tree. It
11
CHAPTER 7
turns out that if a matrix is additive, then there exists a unique simple tree fitting this ma-
trix. In the Distance-Based Phylogeny Problem, we will therefore use the terminology
T REE ( D ) to denote the simple tree fitting the additive distance matrix D. Our question,
then, is how to construct T REE ( D ) from D.
EXERCISE BREAK: Prove that every simple tree with n leaves has at most n 2
internal nodes.
Toward An Algorithm for Distance-Based Phylogeny Construction
A quest for neighboring leaves
A natural first step for solving the Distance-Based Phylogeny Problem would be to
ensure that the two closest species with respect to the distance matrix D correspond to
neighbors in T REE ( D ). In other words, the minimum value Di,j should correspond to
leaves i and j having the same parent. In the rest of this chapter, when we refer to the
minimum element of a matrix, we are referring to a minimum off-diagonal element,
i.e., a value Di,j such that i 6= j.
Theorem. Every simple tree with at least three nodes has a pair of neighboring leaves.
Proof. Given a simple tree T with at least three nodes, consider a path P = (v1 , . . . , vk )
that has the maximum number of nodes of any path in T. Because T has at least three
nodes, k must be at least 3. Furthermore, nodes v1 and vk must be leaves, since otherwise
we could extend P into a longer path. Because T is simple, node v2 , which is the parent
of v1 , must have at least three adjacent nodes: v1 , v3 , and yet another node w.
We claim that w is a leaf, which would imply that leaves v1 and w are neighbors. We
will proceed by contradiction: if w were not a leaf, then since T is simple, w would be
adjacent to another node u. As a result, we could form the path P0 = (u, w, v2 , v3 , . . . , vk ),
which contains k + 1 nodes and contradicts our original assumption that P has the
maximum number of nodes. Thus, w must be a leaf, implying that v1 and w are
neighbors.
Figure 7.8 (top) illustrates that for neighboring leaves i and j sharing a parent node
m, the following equality holds for every other leaf k in the tree:
(di,m + dk,m ) + (d j,m + dk,m ) (di,m + d j,m ) di,k + d j,k di,j

dk,m = = .
2 2
12
i di, k = di, m + dk, m k
di, j = di, m + dj, m m
j dj, k = dj, m + dk, m
i k
di, j = di, m + dj, m m
FIGURE 7.8 For neighboring leaves i and j and their parent node m, dk,m = (di,k +
dj,k di,j )/2 for every other leaf k in the tree. (Bottom) Removing leaves i and j from
the tree turns m into a leaf (we assume that m has degree 3). The distances from this
new leaf to any other leaf k can be recomputed as dk,m = (Di,k + Dj,k Di, j )/2.
Since i, j, and k are leaves, we can compute the distance dk,m between nodes k and m
in terms of elements of the additive distance matrix D,
( Di,k + D j,k Di,j )
dk,m = .
2
In the case when the parent m has degree 3 (as in Figure 7.8 (top)), removing leaves
i and j from the tree turns m into a leaf and thus reduces the total number of leaves
(Figure 7.8 (bottom)). This operation is equivalent to removing rows i and j as well as
columns i and j from D, then adding a new row and column corresponding to their
parent m, where the distances from m to other leaves are computed according to the
above formula.
EXERCISE BREAK: We have just described how to reduce the size of the tree as
well as the dimension of the distance matrix D if the parent node (m) has degree
3. Design a similar approach in the case that the degree of m is larger than 3.
This discussion implies a recursive algorithm for the Distance-Based Phylogeny Prob-
lem:
• find a pair of neighboring leaves i and j by selecting the minimum element Di,j in
the distance matrix;
13
CHAPTER 7
• replace i and j with their parent, and recompute the distances from this parent to
all other leaves as described above;
• solve the Distance-Based Phylogeny problem for the smaller tree;
• add the previously removed leaves i and j back to the tree.
EXERCISE BREAK: Apply this recursive approach to the distance matrix shown
in Figure 7.9 (left). (Solve this exercise by hand.)
i j k l i j k l
i 0 13 21 22 i 0 3 4 3
j 13 0 12 13 j 3 0 4 5
k 21 12 0 13 k 4 4 0 2
l 22 13 13 0 l 3 5 2 0
FIGURE 7.9 (Left) An additive 4 ⇥ 4 distance matrix. (Right) A non-additive 4 ⇥ 4
distance matrix.
Computing limb lengths
If you attempted the preceding exercise, then you were likely driven crazy. The reason
why is that in the first step of our proposed algorithm, we assumed that a minimum
element of an additive distance matrix corresponds to neighboring leaves. Yet as
illustrated in Figure 7.10, this assumption is not necessarily true! Thus, we need a new
approach to the Distance-Based Phylogeny Problem, as finding the animal coronavirus
that is the smallest distance from SARS-CoV may not be the best way to identify the
animal reservoir of SARS.
Our proposed recursive approach may have failed, but using recursion was a good
idea, and so we will explore a different recursive algorithm. Rather than looking for a
pair of neighbors in T REE ( D ), we will instead reduce the size of the tree by trimming its
leaves one at a time. Of course, we don’t know T REE ( D ), and so we must somehow trim
leaves in T REE ( D ) by analyzing the distance matrix.
As a first step toward constructing T REE ( D ), we will address the more modest goal
of computing the lengths of limbs in T REE ( D ). So, given a leaf j in a tree, we denote the
length of the limb connecting j with its parent as L IMB L ENGTH ( j). Edges that are not
limbs must connect two internal nodes and are therefore called internal edges.
14
i k
11 6
4
2 7
j l
FIGURE 7.10 The simple tree fitting the distance matrix from Figure 7.9 (left). The two
closest leaves in this tree (j and k) are not neighbors.
Limb Length Problem:

Compute the length of a limb in the simple tree fitting an additive distance matrix.
Input: An additive distance matrix D and an integer j.

Output: L IMB L ENGTH ( j), the length of the limb connecting leaf j to its
parent in T REE ( D ).
To compute L IMB L ENGTH ( j) for a given leaf j, note that because T REE ( D ) is simple, we
know that PARENT ( j) has degree at least 3 (unless T REE ( D ) has only two nodes). We
can therefore think of PARENT ( j) as dividing the other nodes of T REE ( D ) into at least
three subtrees, or smaller trees that would remain if we were to remove PARENT ( j)
along with any edges connecting it to other nodes (Figure 7.11). Because j is a leaf, it
must belong to a subtree by itself; we call this subtree Tj . This brings us to the following
result.
Limb Length Theorem: Given an additive matrix D and a leaf j, L IMB L ENGTH ( j) is equal
to the minimum value of ( Di,j + D j,k Di,k )/2 over all leaves i and k.
Proof. A given pair of leaves can belong to the same subtree or to different subtrees.
So first assume that leaves i and k belong to different subtrees Ti and Tk (Figure 7.11).
Because PARENT ( j) is on the path connecting i to k, it follows that
di,j = di, PARENT( j) + L IMB L ENGTH ( j)

d j,k = dk, PARENT( j) + L IMB L ENGTH ( j)
Adding these two equations yields
di,j + d j,k = di, PARENT( j) + dk, PARENT( j) + 2 · L IMB L ENGTH ( j).
15
CHAPTER 7
Because di, PARENT( j) + dk, PARENT( j) is equal to di,k , it follows that

di,j + d j,k di,k Di,j + D j,k Di,k
L IMB L ENGTH ( j) = = .
2 2
i
Ti
k
Tk
l
Tj
j
FIGURE 7.11 A simple tree with selected leaves i, j, k, and l. Removing the parent
of j (along with the three dashed edges connecting it to other nodes) would separate
this tree into three subtrees, whose edges are shown in different colors. Leaves i and l
belong to Ti , whereas leaf k belongs to Tk . Leaf j belongs to Tj , which contains a single
node.
On the other hand, assume that leaves i and l belong to the same subtree (Figure 7.11).
Then the path from i to l does not pass through PARENT ( j), and so we have the inequality
di, PARENT( j) + dl, PARENT( j) di,l .
Combining this with the equation
di,j + d j, l = di,PARENT( j) + dl,PARENT( j) + 2 · L IMB L ENGTH ( j)
yields that
di,j + d j,l (di, PARENT( j) + dl, PARENT( j) )

L IMB L ENGTH ( j) =
2
di,j + d j,l di,l Di,j + D j,l Di,l
 = .
2 2
As a result of this discussion, L IMB L ENGTH ( j) must be less than or equal to ( Di,j +
D j,k Di,k )/2 for any choice of leaves i and k. Because we can always find leaves i and
k belonging to different subtrees (why?), it follows that L IMB L ENGTH ( j) is equal to the
minimum value of ( Di,j + D j,k Di,k )/2 over all choices of i and k.
16
We now have an algorithm for solving the Limb Length Problem. For each j, we can
compute L IMB L ENGTH ( j) by finding the minimum value of ( Di,j + D j,k Di,k )/2 over 7B
all pairs of leaves i and k.
EXERCISE BREAK: The proposed algorithm computes L IMB L ENGTH ( j) in

O n2 time (for an n ⇥ n distance matrix). Design an algorithm that computes
L IMB L ENGTH ( j) in O(n) time.
Additive Phylogeny
Trimming the tree
Since we now know how to find the length of any limb in T REE ( D ), we can construct
T REE ( D ) recursively using the algorithm illustrated in Figure 7.12.
First, imagine that we already know T REE ( D ), and pick an arbitrary leaf j. We will
trim the limb of j by reducing its length by L IMB L ENGTH ( j). Because we do not know
T REE ( D ), we need to represent trimming the leaf j in terms of the distance matrix D.
To do so, we first subtract L IMB L ENGTH ( j) from each off-diagonal element in row j
and column j of D to obtain a matrix Dbald for which the limb of j has become a bald
limb, or a limb of length 0 (Figure 7.12). We will further assume that a bald limb has
disappeared from the tree entirely. In terms of the distance matrix, ignoring a bald limb
means removing row j and column j from D to produce a smaller (n 1) ⇥ (n 1)
distance matrix Dtrimmed . We can now recursively find T REE ( D ) in four steps:
• pick an arbitrary leaf j, compute L IMB L ENGTH ( j), and construct the distance
matrix Dtrimmed ;
• solve the Distance-Based Phylogeny Problem for Dtrimmed ;
• identify the point in T REE ( Dtrimmed ) where leaf j should be attached in T REE ( D );
• add a limb of length L IMB L ENGTH ( j) growing from this attachment point in
T REE ( Dtrimmed ) to form T REE ( D ).
STOP and Think: When adding leaf j back to T REE ( Dtrimmed ), how would you
find its attachment point?
17
CHAPTER 7
i j k l
i 0 13 21 22 i k
11 6
j 4
D 13 0 12 13
TREE(D)
k 21 12 0 13 2 7
j l
l 22 13 13 0
LIMBLENGTH(j) = 2
i j k l
i 0 11 21 22 i k
11 6
j 11 0 10 11 4
Dbald TREE(Dbald)
k 21 10 0 13 0 7
j l
l 22 11 13 0
i j k l
i 0 11 21 22 k
6
j 11 0 10 11 15
Dtrimmed i TREE(Dtrimmed)
k 21 10 0 13 7
l
l 22 11 13 0
LIMBLENGTH(l) = 7
i j k l
i 0 11 21 15 k
6
15
j 11 0 10 4 i
k 21 10 0 6 0
l
l 15 4 6 0
i j k l
i 0 11 21 15
j 11 0 10 4 21
i k
k 21 10 0 6
l 15 4 6 0
FIGURE 7.12 Converting the additive distance matrix from Figure 7.9 (left) into the
simple tree fitting this matrix from Figure 7.10. On the left side, we first compute
LIMBLENGTH (j) = 2, and then subtract 2 from the non-diagonal elements in row j and
column j of D to obtain Dbald (updated values are shown in red). Removing this row
and column yields a 3 ⇥ 3 distance matrix Dtrimmed . We find that LIMBLENGTH (l) = 7
in Dtrimmed and subtract 7 from the non-diagonal elements in row l and column l.
Graying out this row and column yields a 2 ⇥ 2 distance matrix. On the right side, we
can fit this 2 ⇥ 2 distance matrix to a tree consisting of a single edge. By finding the
attachment points of removed limbs (shown on the left), we reconstruct TREE (Dtrimmed ),
TREE (Dbald ), and then TREE (D).
18
Attaching a limb
To find the attachment point of a leaf j in T REE ( Dtrimmed ), consider T REE ( Dbald ), which
is the same as T REE ( D ) except that L IMB L ENGTH ( j) = 0. From the Limb Length
Theorem, we know that there must be leaves i and k in T REE ( Dbald ) such that
bald + Dbald
Di,j bald
Di,k
j,k
= 0,
2
which implies that
bald = Dbald + Dbald .
Di,k i,j j,k
bald from leaf i

Thus, the attachment point for leaf j must be located at distance Di,j
on the path connecting i and k in the trimmed tree. This attachment point may occur
at an existing node, in which case we connect j to this node. On the other hand, the
attachment point for j may occur along an edge, in which case we place a new node at
the attachment point and connect j to it.
An algorithm for distance-based phylogeny construction
The preceding discussion results in the recursive algorithm below, which we call
A DDITIVE P HYLOGENY, for finding the simple tree fitting an n ⇥ n additive distance
matrix D. We assume that you have already implemented a program L IMB ( D, j) that
computes L IMB L ENGTH ( j) for a leaf j based on the distance matrix D. Rather than
selecting an arbitrary leaf j from T REE ( D ) for trimming, A DDITIVE P HYLOGENY selects
leaf n (corresponding to the last row and column of D).
STOP and Think: Consider these questions about A DDITIVE P HYLOGENY.
• What is its running time?
• Although it may seem that A DDITIVE P HYLOGENY would construct a

tree for any matrix, this is not the case. What goes wrong if you apply
A DDITIVE P HYLOGENY to the non-additive distance matrix in Figure 7.9
(right)?
• Modify A DDITIVE P HYLOGENY to develop an algorithm that checks

whether a given distance matrix is additive. Then, apply this test to the
distance matrix for coronavirus Spike proteins shown in Figure 7.13 (top).
Is this matrix additive?
19
CHAPTER 7
ADDITIVEPHYLOGENY(D, n)
if n = 2
return the tree consisting of a single edge of length D1,2
limbLength LIMB (D, n)
for j 1 to n 1
Dj , n Dj,n limbLength
Dn,j Dj,n
7C
(i, n, k) three leaves such that Di,k = Di,n + Dn,k
x Di,n
remove row n and column n from D
T ADDITIVEPHYLOGENY (D, n 1)
v the (potentially new) node in T at distance x from i on the path between i and k
add leaf n back to T by creating a limb (v, n) of length limbLength
return T
Although the previous question suggests that A DDITIVE P HYLOGENY can be modified
to determine whether a given distance matrix is additive, there exists an even simpler
PAGE 52 way to check for additivity (see DETOUR: The Four Point Condition).
Constructing an evolutionary tree of coronaviruses
By the end of 2003, bioinformaticians had sequenced many coronaviruses from a variety
of birds and mammals, from which we obtain the distance matrix in Figure 7.13 (top)
based on a multiple alignment of Spike proteins.
Although you now understand the perils of concluding that the minimum element
of the distance matrix corresponds to a pair of neighbors, common sense tells us with
a glance at Figure 7.13 (top) that the civet must be the animal reservoir of SARS. This
information led researchers to hypothesize that inadequate preparation of meat from
palm civets (Figure 7.14) in the Guangdong region of China may have caused the SARS
outbreak.
PAGE 54 Yet before rushing to this conclusion, you may like to read DETOUR: Did Bats Give
Us SARS? to see why the history of interspecies viral transfer is often difficult to trace.
In fact, some studies have suggested that humans first received SARS from bats, which
later gave the virus to palm civets, which then transmitted the disease back to humans.
The civet was identified as the animal reservoir of SARS in 2003 in part because SARS
viruses from other potential suspects, including bats, had not yet been sequenced.
20
STOP and Think: It turns out that most distance matrices constructed from real
data (including the distance matrix in Figure 7.13 (top)) are non-additive. Why
do you think that this is the case?
Since the distance matrix for SARS-like coronaviruses is non-additive, we will cheat a bit
and slightly modify it to make it additive so that you can apply A DDITIVE P HYLOGENY
to it (Figure 7.13 (bottom)).
Cow Pig Horse Mouse Dog Cat Turkey Civet Human

Cow 0 295 300 524 1077 1080 978 941 940
Pig 295 0 314 487 1071 1088 1010 963 966
Horse 300 314 0 472 1085 1088 1025 965 956
Mouse 524 487 472 0 1101 1099 1021 962 965
Dog 1076 1070 1085 1101 0 818 1053 1057 1054
Cat 1082 1088 1088 1098 818 0 1070 1085 1080
Turkey 976 1011 1025 1021 1053 1070 0 963 961
Civet 941 963 965 962 1057 1085 963 0 16
Human 940 966 956 965 1054 1080 961 16 0
Cow Pig Horse Mouse Dog Cat Turkey Civet Human

Cow 0 295 306 497 1081 1091 1003 956 954
Pig 295 0 309 500 1084 1094 1006 959 957
Horse 306 309 0 489 1073 1083 995 948 946
Mouse 497 500 489 0 1092 1102 1014 967 965
Dog 1081 1084 1073 1092 0 818 1056 1053 1051
Cat 1091 1094 1083 1102 818 0 1066 1063 1061
Turkey 1003 1006 995 1014 1056 1066 0 975 973
Civet 956 959 948 967 1053 1063 975 0 16
Human 954 957 946 965 1051 1061 973 16 0
FIGURE 7.13 (Top) The distance matrix based on pairwise alignment of Spike proteins
from coronaviruses extracted from various animals. The distance between each pair of
sequences was computed as the total number of mismatches and indels in their optimal
alignment. (Bottom) A modification of the distance matrix to make it additive.
EXERCISE BREAK: Construct the simple tree fitting the distance matrix in
Figure 7.13 (bottom).
21
CHAPTER 7
FIGURE 7.14 The palm civet.
Using Least Squares to Construct Approximate Distance-Based Phylogenies
If an n ⇥ n distance matrix D is non-additive, then we will instead look for a weighted

tree T whose distances between leaves approximate the entries in D. To this end, we
would like for T to minimize the sum of squared errors D ISCREPANCY ( T, D ), which is
given by the formula
D ISCREPANCY ( T, D ) = Â1i< jn (di,j ( T ) Di,j )2 .
Least Squares Distance-Based Phylogeny Problem:

Given a distance matrix, find the tree that minimizes the sum of squared errors.
Input: An n ⇥ n distance matrix D.

Output: A weighted tree T minimizing D ISCREPANCY ( T, D ) over all
weighted trees with n leaves.
EXERCISE BREAK: Let T be the tree in Figure 7.10 with all edge lengths re-
moved. Given the non-additive 4 ⇥ 4 distance matrix D in Figure 7.9 (right), find
the lengths of edges in this tree that minimize D ISCREPANCY ( T, D ).
It turns out that for a specific tree T, it is easy to find edge weights in T minimizing
D ISCREPANCY ( T, D ). Yet our ability to minimize the sum of squared errors for a specific
tree does not imply that we can efficiently solve the Least Squares Distance-Based Phy-
logeny Problem, since the number of different trees grows very quickly as the number
of leaves in the tree increases. In fact, the Least Squares Distance-Based Phylogeny
winds up being NP-Complete, and so we must abandon the hope of designing a fast
22
algorithm to find a tree that best fits a non-additive matrix. In the next two sections, we
will explore heuristics for constructing trees from non-additive matrices that solve this
problem approximately.
Ultrametric Evolutionary Trees
Biologists often assume that every internal node in an evolutionary tree corresponds to
a species that underwent a speciation event, splitting one ancestral species into two de-
scendants. Note that every internal node in the tree in Figure 7.15 (top) (corresponding
to a speciation event) has degree 3. We therefore define an unrooted binary tree as a
tree where every node has degree equal to either 1 or 3.
EXERCISE BREAK: Prove that every unrooted binary tree with n leaves has
n 2 internal nodes (and thus 2n 3 edges).
A rooted binary tree is an unrooted binary tree that has a root (of degree 2) placed on
one of its edges. In other words, we replace an edge (v, w) with a root and draw edges
connecting the root to each of v and w (Figure 7.15 (bottom)).
If we had a molecular clock measuring evolutionary time, then we could assign an
age to every node v in a rooted binary tree (denoted A GE (v)), where all of the leaves
of the tree have age 0 because they correspond to present-day species. We could then
define the weight of an edge (v, w) in the tree as the difference A GE (v) A GE (w).
Consequently, the length of a path between the root and any node would be equal to
the difference between their ages. Such a tree, in which the distance from the root to
any leaf is the same, is called ultrametric (Figure 7.16 (bottom right)).
Our aim is to derive an ultrametric tree that explains a given distance matrix (even
if it does so only approximately). UPGMA (which stands for Unweighted Pair Group
Method with Arithmetic Mean) is a simple clustering heuristic that introduces a hypo-
thetical molecular clock for constructing an ultrametric evolutionary tree. You can learn
more about clustering in Chapter 8.
Given an n ⇥ n matrix D, UPGMA (which is illustrated in Figure 7.16) first forms n
trivial clusters, each containing a single leaf. The algorithm then finds a pair of “closest”
clusters. To clarify the notion of closest clusters, UPGMA defines the distance between
clusters C1 and C2 as the average pairwise distance between elements of C1 and C2 ,
Â i2C1 Â j2C2 Di,j

DC1 ,C2 = .
|C1 | · |C2 |
23
CHAPTER 7
Chimpanzee
Orangutan
Squirrel
Monkey
Bonobo
Baboon Human
Gorilla
33
23
13
Squirrel Baboon Orangutan Gorilla Chimpanzee Bonobo Human

Monkey
FIGURE 7.15 (Top) An unrooted binary tree representing a phylogeny of primates.

(Bottom) Placing a root on the squirrel monkey’s limb results in a rooted binary tree.
The number at each node corresponds to the number of million years ago that the
divergence at this node occurred.
24
In this equation, the notation |C | denotes the number of leaves in cluster C.

Once UPGMA has identified a pair of closest clusters C1 and C2 , it merges them
into a cluster C with |C1 | + |C2 | elements and then creates a node for C, which it connects
to each of C1 and C2 by a directed edge. The age of C is set to be DC1 ,C2 /2. UPGMA
then iterates this process of merging the two closest clusters until only a single cluster
remains, which corresponds to the root.
UPGMA(D, n)
Clusters n single-element clusters labeled 1, . . . , n
construct a graph T with n isolated nodes labeled by single elements 1, . . . , n
for every node v in T
AGE (v) 0
while there is more than one cluster
find the two closest clusters Ci and Cj (break ties arbitrarily)
merge Ci and Cj into a new cluster Cnew with |Ci | + |Cj | elements
add a new node labeled by cluster Cnew to T
connect node Cnew to Ci and Cj by directed edges 7D
AGE (C) DCi ,Cj /2
remove the rows and columns of D corresponding to Ci and Cj
remove Ci and Cj from Clusters
add a row/column to D for Cnew by computing D(Cnew , C) for each C in Clusters
add Cnew to Clusters
root the node in T corresponding to the remaining cluster
for each edge (v, w) in T
length of (v, w) AGE (v ) AGE (w )
return T
EXERCISE BREAK: Prove that after merging clusters Ci and Cj into a cluster
Cnew , the distance between Cnew and another cluster Cm is equal to ( DCi , Cm ·
|Ci | + DCj , Cm · |Cj |)/ |Ci | + |Cj | .
UPGMA offers a step forward from A DDITIVE P HYLOGENY, since it can analyze non-
additive distance matrices. Figure 7.17 shows the result of applying UPGMA to the coro-
navirus distance matrix from Figure 7.13 (top). However, the first step that UPGMA
takes is to merge the two leaves i and j with minimum distance Di,j into a single clus-
ter. And we have already seen that the smallest element in the distance matrix does
25
CHAPTER 7
i j k l
i 0 3 4 3
j 3 0 4 5
i 0 j 0 k 0 l 0
k 4 4 0 2
l 3 5 2 0
i j k l
i 0 3 4 3 1
j 3 0 4 5 1 1
k 4 4 0 2 i 0 j 0 k 0 l 0
l 3 5 2 0
1.5
i j { k, l }
1
i 0 3 3.5 1.5 1.5
1 1
j 3 0 4.5
{ k, l } 3.5 4.5 0 i 0 j 0 k 0 l 0
2
0.5
1
{i, j} { k, l } 1.5
{i, j} 0 4 1
{ k, l } 4 0 1.5 1.5
1 1
i 0 j 0 k 0 l 0
FIGURE 7.16 Tree reconstruction with UPGMA for the non-additive distance matrix
from Figure 7.9 (right). UPGMA begins with forming one cluster for each leaf. In each
step, it identifies the two closest clusters C1 and C2 , merge them into a new node C,
and connect C to C1 and C2 by directed edges. The age of C is set equal to DC1 ,C2 /2.
We then iterate this process until only a single cluster remains, which must be the root.
The resulting tree is ultrametric (i.e., the distance from the root to any leaf is the same).
26
not necessarily correspond to a pair of neighboring leaves! This is a concern, since

if UPGMA generates incorrect trees from additive matrices, then it is not an ideal
heuristic for evolutionary tree construction from non-additive matrices. Can we find an
algorithm that always identifies neighboring leaves in an additive distance matrix but
also performs well on a non-additive distance matrix?
Cow
147.5
Pig
153.5
247.16 Horse
Mouse
478.62
Palm Civet
8
496.41 Human
538.82 Turkey
Dog
409
Cat
FIGURE 7.17 The ultrametric tree of coronaviruses created by UPGMA using the dis-
tance matrix in Figure 7.13 (top). The root is shown in green.
The Neighbor-Joining Algorithm
Transforming a distance matrix into a neighbor-joining matrix
In 1987, Naruya Saitou and Masatoshi Nei developed the neighbor-joining algorithm
for evolutionary tree reconstruction. Given an additive distance matrix, this algorithm,
which we call N EIGHBOR J OINING, finds a pair of neighboring leaves and substitutes
them by a single leaf, thus reducing the size of the tree. N EIGHBOR J OINING can thus
recursively construct a tree fitting the additive matrix. This algorithm also provides a
heuristic for non-additive distance matrices that performs well in practice.
The central idea of N EIGHBOR J OINING is that although finding a minimum element
in a distance matrix D is not guaranteed to yield a pair of neighbors in T REE ( D ), we
can transform D into a different matrix whose minimum element does yield a pair of
27
CHAPTER 7
neighbors. First, given an n ⇥ n distance matrix D, we define T OTAL D ISTANCE D (i ) as

the sum Â1kn Di,k of distances from leaf i to all other leaves. The neighbor-joining
matrix D ⇤ is defined such that for all i and j, Di,i
⇤ = 0 and
⇤ = (n
Di,j 2) · Di,j T OTAL D ISTANCE D (i ) T OTAL D ISTANCE D ( j) .
N EIGHBOR J OINING, which is illustrated in Figure 7.18, is a widely used method

for evolutionary tree reconstruction; the paper that introduced it is one of the most
cited in all of science, with over 30,000 citations. Yet this algorithm is non-intuitive: the
above formula for computing the matrix D ⇤ probably looks like witchcraft to you. In
fact, despite having flawless intuition, Saitou and Nei never proved that their algorithm
correctly solves the Distance-Based Phylogeny Problem for additive matrices! It took
researchers another decade to prove the following theorem, whose proof we have
PAGE 56 passed to DETOUR: Why Does the Neighbor-Joining Algorithm Find Neighboring
Leaves?
Neighbor-Joining Theorem: Given an additive matrix D, the smallest element Di,j ⇤ of its
neighbor-joining matrix D ⇤ corresponds to a pair of neighboring leaves i and j in T REE ( D ).
If n = 2, then N EIGHBOR J OINING(D, n) returns the tree consisting of a single edge

of length D1,2 . If n > 2, then it selects the minimum element in the neighbor-joining
matrix, replaces the neighboring leaves i and j with a new leaf m, and then computes
the distance from m to any other leaf k according to the formula
Dk,m = 12 ( Dk,i + Dk,j Di,j ),
which is motivated by Figure 7.8. This equation allows us to replace an n ⇥ n matrix

D with an (n 1) ⇥ (n 1) matrix D 0 in which i and j have been replaced by m. By
recursively applying N EIGHBOR J OINING to D 0 , we obtain an evolutionary tree with
n 1 leaves. We then add two limbs starting at node m, one ending in leaf i and the
other ending in leaf j. We set
T OTAL D ISTANCE D (i ) T OTAL D ISTANCE D ( j)

Di,j =
n 2
and assign
1
L IMB L ENGTH (i ) = Di,j + Di,j
2
1
L IMB L ENGTH ( j) = Di,j Di,j
2
28
i j k l TotalDistanceD
i k
i 0 13 21 22 56 11 6
j 13 0 12 13 38 4
D
k 21 12 0 13 46
2 7
l 22 13 13 0 48 j l
i j k l
i 0 -68 -60 -60
j -68 0 -60 -60
D*
k -60 -60 0 -68
l -60 -60 -68 0
m k l TotalDistanceD i k
6
m 0 10 11 21
m 4
D k 10 0 13 23
l 11 13 0 24 7
j l
m k l
m 0 -34 -34
D* k -34 0 -34
l -34 -34 0
i k
m a
m 0 4 m 4 a
D a 4 0
j l
FIGURE 7.18 (Top left) The additive distance matrix D from Figure 7.9 (left) along
with the array TotalDistanceD . Dj,k (shown in blue) is a minimum element of D,
but as it turns out, leaves j and k are not neighbors in TREE (D). Moving down the
left side, we construct the neighbor-joining matrix D⇤ and find that D⇤i,j (red) is a
minimum element of D⇤ . We transform the initial 4 ⇥ 4 distance matrix into a 3 ⇥ 3
distance matrix by replacing i and j with a single leaf m and updating distances from
m to other leaves as Dk,m = 12 (Dk,i + Dk,j Di,j ) = 12 (21 + 12 13) = 10 and
Dl,m = 12 (Dl,i + Dl,j Di,j ) = 12 (22 + 13 13) = 11. We select D⇤k,l as a minimum
element of D⇤ and replace leaves k and l with a single leaf a. The resulting 2 ⇥ 2 matrix
corresponds to a tree with a single edge connecting m and a. We then work our way up
the right side, adding pairs of neighbors back into the tree at each step using formulas
for limb lengths. (Top right) The tree TREE (D) fitting the original matrix D.
29
CHAPTER 7
PAGE 61 To see where these formulas come from, see DETOUR: Computing Limb Lengths in
the Neighbor-Joining Algorithm.
EXERCISE BREAK: Prove that for any i and j between 1 and n, both 12 ( Di,j + Di,j )
and 12 ( Di,j Di,j ) are non-negative.
The following pseudocode summarizes the neighbor-joining algorithm.
NEIGHBORJOINING(D, n)
if n = 2
T the tree consisting of a single edge of length D1,2
return T
D⇤ the neighbor-joining matrix constructed from the distance matrix D
find elements i and j such that D⇤i ,j is a minimum non-diagonal element of D⇤
(TOTALDISTANCED (i) TOTALDISTANCED (j))/(n 2)
1
limbLengthi 2 (Di,j + )
1
limbLengthj 2 (Di,j )
7E add a new row/column m to D so that Dk,m = Dm,k = 12 (Dk,i + Dk,j Di,j )
for any k
remove rows i and j from D
remove columns i and j from D
T NEIGHBORJOINING(D, n 1)
add two new limbs (connecting node m with leaves i and j) to the tree T
assign length limbLengthi to LIMB (i)
assign length limbLengthj to LIMB (j)
return T
EXERCISE BREAK: Before implementing N EIGHBOR J OINING, apply it to the

additive and non-additive distance matrices from Figure 7.9.
EXERCISE BREAK: Apply N EIGHBOR J OINING to the coronavirus distance

matrix from Figure 7.13 (top).
30
Analyzing coronaviruses with the neighbor-joining algorithm
Figure 7.19 shows the neighbor-joining tree of coronaviruses isolated from differ-
ent animals based on the distance matrix in Figure 7.13 (top). We can also apply
N EIGHBOR J OINING to the distance matrix of SARS-CoV variants isolated from various
human carriers, in addition to a coronavirus from palm civet (Figure 7.20).
146
Cow
11 149
86 Pig
251 149
Horse
Mouse
254
22
7 Palm Civet
455
9
Human
489
163
Turkey
404
Dog
414
Cat
FIGURE 7.19 The neighbor-joining tree of coronaviruses isolated from different animals,
based on the non-additive distance matrix in Figure 7.13 (top).
The SARS-CoV strain labeled “Hanoi” in Figure 7.20 was taken from Carlo Urbani,
an Italian physician who worked for the World Health Organization. In February 2003,
Urbani was called into a Hanoi hospital to examine a patient who had fallen ill with
what local doctors believed to be a bad case of influenza; these doctors were afraid that
it might be bird flu. In fact, the patient was the American man who had stayed across
the hall from Liu Jianlun at the Metropole Hotel just one week earlier.
Fortunately, Urbani was quick to realize that the disease was not influenza, and
he became the first physician to raise the alarm to public health officials. Yet rather
than leaving Hanoi, he demanded that he remain there in order to oversee quarantine
procedures. In an argument with his wife, who scolded him for risking his life to treat
sick patients. Urbani replied, “What am I here for? Answering e-mails, going to cocktail
parties and pushing paper?” Urbani would ultimately lose his life to SARS a month
31
CHAPTER 7
Guangzhou Zhongshan Guangzhou Guangzhou Guangzhou Hong Kong Hanoi Toronto Hong Kong Palm
Dec. 16, Dec. 16, Jan. 24, Jan. 31, Feb. 18, Feb. 18, Feb. 26, Feb. 17, Mar. 15, Civet
2002 2002 2003 2003 2003 2003 2003 2003 2003
Guangzhou 0 4 12 8 9 9 12 12 11 3
Zhongshan 4 0 10 6 7 7 10 10 9 3
Guangzhou 12 10 0 4 5 3 2 2 1 11
Guangzhou 8 6 4 0 3 1 4 4 3 7
Guangzhou 9 7 5 3 0 2 5 5 4 8
Hong Kong 9 7 3 1 2 0 3 3 2 8
Hanoi 12 10 2 4 5 3 0 2 1 11
Toronto 12 10 2 4 5 3 2 0 1 11
Hong Kong 11 9 1 3 4 2 1 1 0 10
Palm Civet 3 3 11 7 8 8 11 11 10 0
0.17
Guangzhou: January 24, 2003
0
Guangzhou: February 18, 2003
1
Zhongshan: January 4, 2003
1
Hong Kong: February 21, 2003
0.5 2
0
Toronto: February 27, 2003
1
4.83 Hanoi: February 26, 2003
1.5
Guangzhou: January 31, 2003
1
Zhongshan: December 26, 2002
1
Palm Civet
1
2
Guangzhou: December 16, 2002
FIGURE 7.20 (Top) The distance matrix based on pairwise alignment of Spike proteins
from SARS-CoV strains extracted from various patients as well as a coronavirus taken
from a palm civet. The distance between each pair of sequences was computed as the
total number of mismatches and indels in their optimal pairwise alignment. (Bottom)
The evolutionary tree of these viruses constructed by the neighbor-joining algorithm.
32
later. But his sacrifice also helped begin the massive worldwide action against this
disease that may well have saved millions of lives.
Limitations of distance-based approaches to evolutionary tree construction
Although distance-based tree reconstruction has successfully resolved questions about

the origin and spread of SARS, many evolutionary controversies cannot be resolved by
using a distance matrix. For example, when we convert each pair of rows of a multiple
alignment into a distance value, we lose information contained in the alignment. As a
result, distance-based methods do not allow us to reconstruct ancestral sequences of
Spike proteins (corresponding to internal nodes in Figure 7.20), which could lead us to
believe that such molecular paleontology is impossible. Therefore, a superior approach
to evolutionary tree reconstruction would be to somehow use the alignment directly,
without first converting it to a distance matrix.
Character-Based Tree Reconstruction
Character tables
Fifty years ago, biologists constructed phylogenies not from DNA and protein sequences
but from anatomical and physiological features called characters. For example, when
analyzing invertebrate evolution, one commonly used character is the presence or
absence of wings, while another is the number of legs (varying from 0 to over 300 in
some centipedes). These two characters result in the 3 ⇥ 2 character table shown in
Figure 7.21 for three species.
In general, every row in an n ⇥ m character table represents a character vector,
holding the values of m characters corresponding to one of n existing species. Our goal,
roughly stated, is to construct an evolutionary tree in which leaves corresponding to
present-day species with similar character vectors occur near each other in the tree. We
would also like to assign m character values to each internal node in the tree in order to
best explain the characters of ancestral species.
STOP and Think: Can you transform the preceding vague description into a
well-formulated computational problem that models tree reconstruction based
on character tables?
33
CHAPTER 7
wings legs
winged stick insect Yes 6
wingless stick insect No 6
giant centipede No 42
FIGURE 7.21 (Top panel) Winged (left) and wingless (middle) stick insects, each having
six legs, and the giant centipede (right), which has 42 legs. (Bottom panel) A 3 ⇥ 2
character table describes two characters (wings and legs) in these three invertebrates.
From anatomical to genetic characters
In 1965, Emile Zuckerkandl and Linus Pauling published “Molecules as documents

of evolutionary history”, arguing that DNA sequences offer a much more informative
source of data than anatomical and physiological characters. The idea may seem obvious
today, especially after we have spent half of a chapter constructing evolutionary trees
from distance matrices generated from DNA sequences. However, Zuckerkandl and
Pauling’s proposal was initially met with skepticism by many biologists, who felt that
DNA analysis could not offer the same power as anatomical comparison. A now-
famous argument was initiated when Zuckerkandl and Pauling found that the amino
acid sequence of human beta-hemoglobin is very similar to that of gorillas, prompting
Zuckerkandl to write in 1963:
From the point of view of hemoglobin structure, it appears that gorilla is just an
abnormal human.
Yet the surprising similarity between hemoglobin proteins in various primates flew
in the face of clear anatomical differences between primates. As a result, the leading
evolutionary biologist Gaylord Simpson immediately responded to Zuckerkandl:
...that is of course nonsense. What the comparison really indicates is that hemoglobin
is a bad choice and has nothing to tell us about attributes, or indeed tells us a lie.
34
Despite such vehement initial criticisms, genetic analysis had become the dominant
technique in evolutionary studies by the 1970s. Indeed, the analysis of DNA sequences
answered evolutionary questions that previous analysis of anatomical characters had
failed to resolve. Early examples include the classification of the giant panda (see
DETOUR: Giant Panda: Bear or Raccoon?) and the identification of human origins PAGE 62
(see DETOUR: Where Did Humans Come From?). As a result, most evolutionary
experts had no choice but to adapt their views and become experts on molecular PAGE 62
evolution.
Ironically, modern evolutionary studies often view an n ⇥ m multiple alignment as
an n ⇥ m character table, with each column representing a character of its own. Our
goal is to construct a tree whose leaves correspond to the rows of this alignment and
whose internal nodes correspond to ancestral sequences in accordance with the most
parsimonious evolutionary scenario. Before rigorously defining a most parsimonious
scenario, we will describe one example of how this algorithmic framework solved a
longstanding puzzle in insect evolution.
How many times has evolution invented insect wings?
Wings provided a revolutionary adaptation for insects, allowing them to escape preda-
tors and disperse into new territories, thus leading to many new insect species. Yet
despite the evolutionary advantages provided by wings, some insects are apparently
better equipped for survival without them. In fact, nearly all winged species are related
to many wingless counterparts belonging to the same genus, and some entire orders of
insects are wingless, including fleas and lice.
The acquisition of wings would seem to pose an evolutionary challenge because
complex physiological interactions are required to accommodate flight. As a result,
we would be led to believe that wings evolved only once in insects. This argument
parallels Dollo’s principle of irreversibility, a hypothesis proposed by 19th Century
paleontologist Louis Dollo. According to this principle, when a species loses a complex
organ, such as wings, the organ will not reappear in exactly the same form in the
species’s descendants.
STOP and Think: What do you think about this argument as it pertains to insect
wings?
Until recently, biologists followed Dollo’s principle with respect to insect wings, be-
lieving that re-evolution of wings was essentially impossible because unused flight
35
CHAPTER 7
genes in wingless insects would be free to accumulate mutations, eventually eroding

into non-functional pseudogenes. However, in 2003, Michael Whiting studied various
winged and wingless stick insects from around the world and refuted this argument.
He sequenced an approximately 2,000 nucleotide-long segment (the 18S ribosomal
RNA gene) from these stick insects and constructed an evolutionary tree based on these
sequences. From this phylogeny, he inferred that wings were re-invented at least three
times and lost at least four times during stick insect evolution (Figure 7.22).
Oligotoma nigra
Teratembia
Timema knulii
Oreophoetes peruana
Libethra regularis
Diapheromera
Diapheromera femorata
Baculum thaii
Gratidia fritzschei
Sceptrophasma hispidula
S. langikawicensis
Baculini
Medauroidea extradentum
Medaura
Oncotophasma martini
Bacillus rossius
Carausius morosus
Eurycanthia insularis
Phyllium bioculatum
Neohirasea maerens
Neohirasea
Sipyloidea sipylus
Pseudodiacantha macklottii
Lopaphus perakensis
Lopaphus sphalerus
Phoebaticus heusii
Lamponius guerini
Dimorphodes prostasis
Tropiderus childrenii
Eurycnema goliath
Ctenomorphodes briareus
Extatosoma tiaratum
Anisomorpha ferruginea
Pseudophasma
Aretaon asperrimus
Sungaya inexpectata
Agathemera crassa
Heteropteryx dilatata
Haaniella dehaanii
FIGURE 7.22 Evolutionary tree of winged (blue) and wingless (red) stick insects con-
structed from 18S ribosomal RNA genes. Transitions from winged to wingless species are
shown by yellow stars; transitions from wingless to winged species are shown by green
stars. 18S ribosomal RNAs are slow-evolving, making them ideal for reconstructing
ancient divergences.
36
STOP and Think: Is it possible to infer an evolutionary scenario from the tree in
Figure 7.22 that re-invents wings fewer than four times?
Whiting’s work indicates the inherent complications of trying to infer evolutionary trees
from anatomical characters. If you were asked to design a character-based phylogeny
algorithm for the collection of all stick insects, then one of the first steps you would
probably take would be to cluster all the winged insects on the opposite side of the
tree from all of the wingless insects. However, the re-evolution of insect wings means
that such an approach is flawed. Anatomical characters have even less power when we
move into the microscopic world: just imagine trying to construct a phylogeny from
direct observation of coronaviruses!
The Small Parsimony Problem
We will label each leaf of a tree by a row of a multiple alignment, and we will attempt to
infer strings labeling internal nodes and corresponding to candidate ancestral sequences.
However, we will need to develop a scoring function in order to quantify how well
such a labeled tree fits the given multiple alignment. In what follows, we assume for
simplicity that the multiple alignment contains only substitutions and no indels. In
practice, researchers may start from a multiple alignment containing indels and then
remove all columns containing indels.
An intuitive score of an evolutionary tree is the total number of mutations required
to explain the strings at all nodes of the tree. Given a tree T with every node labeled by a
string of length m, we will therefore set the length of edge (v, w) equal to the number of
substitutions (Hamming distance) between the strings labeling v and w. The parsimony
score of T is the sum of the lengths of its edges (Figure 7.23).
We will first assume that we are given the structure of a rooted binary tree in ad-
vance, in which case we only need to assign strings to the internal nodes in order to
minimize the parsimony score.
Small Parsimony Problem:

Find the most parsimonious labeling of the internal nodes of a rooted tree.
Input: A rooted binary tree with each leaf labeled by a string of length m.
Output: A labeling of all other nodes of the tree by strings of length m that
minimizes the tree’s parsimony score.
37
CHAPTER 7
ACGAAAGCCT
1 2
ACGTAAGCCT TCGAAAGCAT
1 2 2 0
ACGTAGGCCT ATGTAAGACT TCGAGAGCAC TCGAAAGCAT

Chimp Human Seal Whale
FIGURE 7.23 An evolutionary tree with parsimony score 8 whose leaves are the DNA
strings in the multiple alignment from Figure 7.3. Colored letters indicate mismatches
in strings connected by an edge.
Given a tree T for which every node v is labeled by a string of length m, we define
labeled trees T1 , . . . , Tm , where Ti has the same structure as T, and where a given node is
labeled by the i-th symbol of the corresponding node in T. Since the parsimony score of
T is the sum of parsimony scores of the trees T1 , . . . , Tm , the Small Parsimony Problem
can be solved independently for each column of the alignment. This observation lets us
assume that every leaf is labeled by a single symbol rather than by a string. Thus, the
weight of an edge connecting two nodes should be either 0 or 1, depending on whether
these nodes are labeled by the same symbol or different symbols; given symbols i and j,
we define di,j = 0 if i = j and di,j = 1 if i 6= j.
We will describe a dynamic programming algorithm, called S MALL PARSIMONY,
for solving the “single character” version of the Small Parsimony Problem. Recall that a
rooted tree T can be viewed as a directed tree with all of its edges directed away from
the root toward the leaves. Thus, every node v in T defines a subtree Tv formed by the
nodes “beneath” v and consisting of all the nodes that can be reached by moving down
from v (Figure 7.24).
Let k be a symbol in a given alphabet and v be a node in a tree T. Define sk (v) as the
minimum parsimony score of the subtree Tv over all possible labelings of the nodes of
Tv such that v is labeled by k. The initial conditions for S MALL PARSIMONY must assign
scores to leaves. If leaf v is labeled by k, then the only character we are allowed to assign
to this leaf is k. Therefore, sk (v) = 0 if leaf v is labeled by symbol k, and si (v) = •
otherwise (Figure 7.25).
38
FIGURE 7.24 The (blue) subtree Tv of a node v within a larger rooted binary tree T.
C C A C G G T C
A C G T A C G T A C G T A C G T A C G T A C G T A C G T A C G T
0 ! 0 ! 0 ! 0 ! !0 !0 ! 0 0 ! !
FIGURE 7.25 Initializing values sk (v) for all leaves v, represented above as an array for
each leaf. We set sk (v) equal to zero if the leaf is labeled by symbol k; otherwise, we
set sk (v) equal to infinity.
If v is an internal node of T, then v is connected to two “children” nodes (the nodes

beneath v in T) that we arbitrarily denote as D AUGHTER (v) and S ON (v). The score
sk (v) can be computed as the minimum of si ( D AUGHTER (v)) + di,k over all possible
symbols i, plus the minimum of s j ( S ON (v)) + dj,k over all possible symbols j:
sk (v) = min {si ( D AUGHTER (v)) + di, k } + min {s j ( S ON (v)) + dj, k }

all symbols i all symbols j
39
CHAPTER 7
This equation allows us to compute all values sk (v) by working our way upward
in T from the leaves to the root, denoted root (Figure 7.26). The subtree of the root is
the entire tree T, and so the minimum parsimony score is given by the smallest score
sk (root) over all symbols k,
min sk (root).
all symbols k
The pseudocode for S MALL PARSIMONY is shown below. It returns the parsimony
score for a rooted binary tree T whose leaves are labeled by symbols stored in an array
C HARACTER (i.e., C HARACTER (v) is the label of leaf v). At each iteration, it selects
a node v and computes sk (v) for each symbol k in the alphabet. For each node v,
S MALL PARSIMONY maintains a value TAG (v), which indicates whether the node has
been processed (i.e., TAG (v) = 1 if the array sk (v) has been computed and TAG (v) = 0
otherwise). We call an internal node of T ripe if its tag is 0 but its children’s tags are
both 1. S MALL PARSIMONY works upward from the leaves, finding a ripe node v at
which to compute sk (v) at each step.
SMALLPARSIMONY(T, CHARACTER)
for each node v in tree T
TAG (v) 0
if v is a leaf
TAG (v ) 1
for each symbol k in the alphabet
if CHARACTER (v) = k
sk ( v ) 0
7F else
sk (v) 1
while there exist ripe nodes in T
v a ripe node in T
TAG (v) 1
for each symbol k in the alphabet
sk ( v ) min {si (DAUGHTER (v)) + di,k } + min {sj (SON (v)) + dj,k }
all symbols i all symbols j
return min sk ( v )
all symbols k
STOP and Think: What is the final ripe node processed by S MALL PARSIMONY,
regardless of the order in which we process ripe nodes?
40
A C G T A C G T A C G T A C G T
2 0 2 2 1 1 2 2 2 2 0 2 2 1 2 1
C C A C G G T C
0 ! 0 ! 0 ! 0 ! !0 !0 ! 0 0 ! !
A C G T A C G T
2 1 3 3 3 2 2 2
2 0 2 2 1 1 2 2 2 2 0 2 2 1 2 1
C C A C G G T C
0 ! 0 ! 0 ! 0 ! !0 !0 ! 0 0 ! !
A C G T
5 3 4 4
C
A C G T A C G T
2 1 3 3 3 2 2 2
2 0 2 2 1 1 2 2 2 2 0 2 2 1 2 1
C C A C G G T C
0 ! 0 ! 0 ! 0 ! !0 !0 ! 0 0 ! !
FIGURE 7.26 An illustration of SMALLPARSIMONY after initialization in Figure 7.25. The

parsimony score is equal to the minimum score at the root, which for this tree is equal
to 3. This value corresponds to symbol C, and so when we begin backtracking to assign
symbols to internal nodes, we assign nucleotide C to the root.
41
CHAPTER 7
Once we compute the parsimony score of a tree T, we also need some way of assigning
symbols to internal nodes of T. The minimum value of sk (root) in Figure 7.26 (bottom) is
equal to 3, which is achieved when k = C. Assigning symbols to the remaining internal
nodes is similar to the backtracking approach that we used for sequence alignment.
Because you are already a pro at dynamic programming, we leave this task to you as an
exercise.
EXERCISE BREAK: Assign nucleotides to all nodes of the tree in Figure 7.26 in
order to solve the Small Parsimony Problem.
When the position of the root in the tree is unknown, we can simply assign the root to
any edge that we like, apply S MALL PARSIMONY to the resulting tree, and then remove
the root. It can be shown that this method provides a solution to the following problem.
Small Parsimony in an Unrooted Tree Problem:

Find the most parsimonious labeling of the internal nodes of an unrooted tree.
Input: An unrooted binary tree with each leaf labeled by a string of length
7G
m.
Output: A labeling of all other nodes of the tree by strings of length m that
minimizes the tree’s parsimony score.
EXERCISE BREAK: Estimate the runtime of the proposed algorithm for solving
the Small Parsimony in an Unrooted Tree Problem.
EXERCISE BREAK: Given a multiple alignment of SARS viruses, reconstruct

the most parsimonious amino acid sequence of the Spike protein in the ancestral
SARS virus (the root of the tree) under the assumption that the evolutionary tree
in Figure 7.17 is correct.
42
The Large Parsimony Problem
S MALL PARSIMONY does not help us if we do not know the evolutionary tree in ad-
vance. In this case, we must find a binary tree as well as assign ancestral strings to
all internal nodes of this tree in order to minimize the parsimony score. Figure 7.27
presents the solutions of the Small Parsimony in an Unrooted Tree Problem for the
three different unrooted binary trees with four leaves, where the leaves are assigned the
strings from the toy multiple alignment in Figure 7.3.
Human Seal
ATGTAAGACT TCGAGAGCAC
2 2
3
ACGTAAGCCT TCGAAAGCAT
1 0
ACGTAGGCCT TCGAAAGCAT
Chimp Whale
Seal Human
TCGAGAGCAC ATGTAAGACT
4 3
0
ACGTAAGCAT ACGTAAGCAT
2 2
ACGTAGGCCT TCGAAAGCAT
Chimp Whale
Whale Human
TCGAAAGCAT ATGTAAGACT
3 2
0
ACGTAAGCCT ACGTAAGCCT
1 5
ACGTAGGCCT TCGAGAGCAC
Chimp Seal
FIGURE 7.27 The three unrooted binary tree structures for the four species from Fig-
ure 7.3, with internal nodes labeled by a solution of the Small Parsimony in an Unrooted
Tree Problem. The first tree solves the Large Parsimony Problem because it has smaller
parsimony score (8) than the other two trees (11 each).
43
CHAPTER 7
Large Parsimony Problem:

Given a set of strings, find a tree — with leaves labeled by all these strings — having
minimum parsimony score.
Input: A collection of strings of equal length.

Output: An unrooted binary tree T that minimizes the parsimony score
among all possible unrooted binary trees with leaves labeled by these strings.
Unfortunately, the Large Parsimony problem is NP-Complete, in part because the num-
ber of different trees grows very quickly with respect to the number of leaves. As a
workaround, we will use a greedy heuristic that explores some but not all trees. First,
note that in any unrooted binary tree, the removal of an internal edge, along with the
two internal nodes that this edge connects, results in four subtrees, which we will
call W, X, Y, and Z (Figure 7.28). These four subtrees can be combined into a tree in
three different ways, which we denote WX |YZ, WY | XZ, and WZ | XY. These three trees
are called nearest neighbors; a nearest neighbor interchange operation replaces a tree
with one of its nearest neighbors.
EXERCISE BREAK: Find the two nearest neighbors of the tree in Figure 7.15 for
the internal edge connecting the parents of gorilla and human.
Like the 2-break operation in Chapter 6, a nearest neighbor interchange corresponds

to replacing two edges in the tree by two new edges. For example, denote the internal
edge of a nearest neighbor interchange as ( a, b); denote the remaining nodes adjacent to
a as w and x; and denote the remaining nodes adjacent to b as y and z. The tree on the
right in Figure 7.28 is obtained from the tree on the left by removing edges ( a, x ) and
(b, y) and replacing them with ( a, y) and (b, x ). The tree on the bottom in Figure 7.28 is
obtained from the tree on the left by removing edges ( a, x ) and (b, z) and substituting
them with ( a, z) and (b, x ).
EXERCISE BREAK: Figure 7.29 shows all possible unrooted binary trees with
five leaves. Find a pair of these trees that are the “farthest apart” in that they
require the maximum number of nearest neighbor interchanges to transform one
tree into the other.
44
W w y Y W w x X
a b a b
X x z Z Y y z Z
W w y Y
a b
Z z x X
FIGURE 7.28 A nearest neighbor interchange on the internal edge (a, b), shown in
black, results from rearranging the four colored subtrees W, X, Y, and Z, which are
rooted at w, x, y, and z, respectively. The nearest neighbor interchange operation
removes one edge connected to a and another edge connected to b, then replaces these
edges with two new edges. The three possible tree structures resulting from nearest
neighbor interchanges on (a, b) can be represented as WX|YZ (top left), WY|XZ (top
right), and WZ|XY (bottom).
Nearest Neighbors of a Tree Problem:

Given an edge in a binary tree, generate the tree’s nearest neighbors.
Input: An internal edge in a binary tree. 7H

Output: The two nearest neighbors of this tree (with respect to the given
internal edge).
The nearest neighbor interchange heuristic for the Large Parsimony Problem starts
from an arbitrary unrooted binary tree. It assigns input strings to arbitrary leaves of
this tree, assigns strings to the internal nodes of the tree by solving the Small Parsimony
Problem in an Unrooted Tree, and then moves to a nearest neighbor that provides the
best improvement in the parsimony score. At each iteration, the algorithm explores all
internal edges of a tree and generates all nearest neighbor interchanges for each internal
edge. For each of these nearest neighbors, the algorithm solves the Small Parsimony
Problem to reconstruct the labels of the internal nodes and computes the parsimony
score. If a nearest neighbor with smaller parsimony score is found, then the algorithm.
selects the one with smallest parsimony score (ties are broken arbitrarily) and iterates
again; otherwise, the algorithm terminates. This is achieved by the following pseu-
docode.
45
CHAPTER 7
j k k j l j
1 2 3
i l i l i k
m m m
j k k j m j
4 5 6
i m i m i k
l l l
j l l j m j
7 8 9
i m i m i l
k k k
k l l k m k
10 11 12
i m i m i l
j j j
k l l k m k
13 14 15
j m j m j l
i i i
FIGURE 7.29 All fifteen unrooted binary trees with five labeled leaves. Tree 1 can be
transformed into trees 4, 7, 12, and 15 by a single nearest neighbor interchange. Note
that every tree has the same structure; this is not the case for trees containing more
than five leaves.
46
NEARESTNEIGHBORINTERCHANGE(Strings)
score 1
generate an arbitrary unrooted binary tree Tree with |Strings| leaves
label the leaves of Tree by arbitrary strings from Strings
solve the Small Parsimony in an Unrooted Tree Problem for Tree
label the internal nodes of Tree according to a most parsimonious labeling
newScore the parsimony score of Tree
newTree Tree
while newScore < score
score newScore 7I
Tree newTree
for each internal edge e in Tree
for each nearest neighbor NeighborTree of Tree with respect to the edge e
solve the Small Parsimony in an Unrooted Tree Problem for NeighborTree
neighborScore the minimum parsimony score of NeighborTree
if neighborScore < newScore
newScore neighborScore
newTree NeighborTree
return newTree
EXERCISE BREAK: Chapter 8 describes how the alcohol dehydrogenase (Adh)

gene helps yeast produce alcohol. To study the evolution of this gene, biologists
constructed a multiple alignment of the Adh genes from various yeast species.
Use this alignment to reconstruct the evolutionary tree of various yeast species
and an ancient yeast Adh ancestor gene.
We have now encountered a number of algorithms for constructing evolutionary trees,

but this does not mean that we can easily resolve various evolutionary controversies.
For example, the identity of the chimpanzee’s closest relative remained undecided until
the mid-1990s (Figure 7.30), and the question of whether mice are closer to humans than
to dogs is still the subject of debate (Figure 7.31).
47
CHAPTER 7
Gorilla Chimpanzee Human Gorilla Chimpanzee Human
FIGURE 7.30 (Left) Analysis of beta-globin genes in human, chimpanzee, and gorilla
suggests a human-chimpanzee split. (Right) Analysis of dopamine D4 receptor gene
suggests a gorilla-chimpanzee split.
Whale
Cow
Dog
Human
Mouse
100 80 60 40 20 0
Millions of years ago
FIGURE 7.31 At the start of the 21st Century, biologists believed that dogs are evolu-
tionarily closer to humans than mice are. However, recent studies suggest otherwise,
as shown in the above phylogeny.
Epilogue: Evolutionary Trees Fight Crime
Janice Trahan met Dr. Richard Schmidt in 1982 when she began working as a nurse in
Lafayette, Louisiana. Both Janice and Richard were married with children, but they
fell in love. Janice soon divorced her husband, but although Richard promised that he
would divorce his wife, he never did. After twelve years, she got tired of waiting, and
broke off the relationship. Two weeks later, she awoke in the middle of the night to see
Richard standing over her, a syringe in his hand.
48
Although Janice had broken up with Richard, she was not surprised to see him. She
had even left her door unlocked for him because he had been giving her vitamin B-12
injections for her chronic fatigue. Since Janice had B-12 injections before, she knew what
to expect. This time, however, she experienced scorching pain as Richard squeezed the
syringe.
A few months later, Janice tested positive for HIV, and she accused Richard of
infecting her via an injection. The police detective in Lafayette had never heard anything
as bizarre as this revenge story — a syringe tainted with HIV had never been used as a
murder weapon. At first, he suspected that Janice had fabricated the story to tarnish the
reputation of her former lover. Nevertheless, he started an investigation by collecting
samples from several HIV patients in Lafayette.
After investigating hospital records, the detective found that Richard had taken
blood from Donald McClelland, an HIV patient, on the same day that he had injected
Janice. Now the forensic challenge was to determine whether the HIV strain taken
from McClelland was similar to Janice’s strain. After HIV DNA was collected from
Janice, McClelland, and many other unrelated HIV-infected patients from Lafayette,
scientists constructed the evolutionary tree of these HIV viruses and found that the
viruses sampled from Janice and McClelland formed a subtree of this tree (Figure 7.32).
The case “State of Louisiana vs. Richard Schmidt” went to trial in 1998. The promi-
nent evolutionary biologist David Hillis presented the evolutionary tree as evidence of
the crime, demonstrating that Janice’s HIV sequence had been derived (with some small
variations) from McClelland’s HIV sequence. Richard Schmidt was then sentenced to
50 years in prison for attempted murder.
STOP and Think: If you had been Richard Schmidt’s attorney, how would you
have argued for his innocence?
CHALLENGE PROBLEM: Given HIV sequences from AIDS patients in

Lafayette, construct the evolutionary tree for other HIV proteins. Does each
tree support conviction of Dr. Schmidt? Reconstruct the ancestral HIV sequences
at the internal nodes of the resulting trees.
49
CHAPTER 7
9.09
JT1
0.97
16.02 JT2
1.63 3.03
14.41 JT3
0.47
25.48 DM1
1.53
DM2
11.88
DM3
22.2
1.26 P1
24.79
P2
0.19 16.83
3.24 P3
23.18
1.11 P4
5.97 20.25
4.88 P5
23.75
3.31 P6
24.62
P7
17.09
P8
1.60
20.91
P9
0.44 15.37
2.46 P10
1.54 1.39 16.63
P11
17.54
P12
0.94 16.36
P13
14.28
P14
FIGURE 7.32 An evolutionary tree of HIV viruses taken from various patients in Lafayette.
Samples from the victim, Janice Trahan (blue leaves JT1, JT2, and JT3), and Richard
Schmidt’s patient, Donald McClelland (red leaves DM1, DM2, DM3), are clustered
together and are rather different than sequences from other patients from Lafayette
(labeled P1 to P14).
50
Detours
When did HIV jump from primates to humans?
Scientists became aware that HIV causes AIDS in the 1980s, at a time when the virus
was still rare. They immediately started looking for earlier cases of HIV infection in
various medical records and found HIV in a blood sample taken from a Congolese
patient in 1959. Genetic studies then revealed that HIV is closely related to Simian
Immunodeficiency Virus (SIV), which infects primates, but it remained unclear how
and when SIV entered the human population and evolved into HIV. The most popular
current hypothesis is that SIV evolved into HIV when hunters killed monkeys to sell
their meat and were exposed to the animals’ blood. The virus circulating in the blood
had further entered cuts in the hunters’ skin, mutated, and later adapted to humans.
By finding regions in the viral genome that mutate at a roughly constant rate
over time, researchers can infer a timeline of HIV evolution. Using this “molecular
clock”, biologists estimated the time points when the various subtypes of SIV jumped
from primates to human: for HIV groups A, B, M, and O, these transitions have been
estimated at 1940, 1945, 1908, and 1920, respectively (the timing of the jump for group
N is currently unknown). By sequencing fecal samples of wild primates, biologists have
even been able to locate populations of chimpanzees and sooty mangabeys whose SIVs
are direct ancestors of HIV groups.
Searching for a tree fitting a distance matrix
Every 3 ⇥ 3 matrix D is additive. To see why, note that there is only one tree with three
leaves. We will denote the leaves of this tree as 1, 2, and 3, and its internal node as c.
As illustrated in Figure 7.33, the lengths of edges in this tree must satisfy the following
three equations:
d1,c + d2,c = D1,2 d1,c + d3,c = D1,3 d2,c + d3,c = D2,3
Solving this system of equations yields the following formulas for the lengths of edges
in terms of values of the matrix D:
D1,2 + D1,3 D2,3 D2,1 + D2,3 D1,3 D3,1 + D3,2 D1,2
d1,c = d2,c = d3,c =
2 2 2
Figure 7.34 illustrates an attempt to fit the distance matrix in Figure 7.9 (right) to
all possible unrooted trees with four leaves. Each such tree leads to a system of six
linear equations (in either four or five variables) that does not have a solution. Thus,
this distance matrix must be non-additive.
51
CHAPTER 7
d3, c
D1, 3 = d1, c + d3, c D2, 3 = d2, c + d3, c
c
d1, c d2, c
1 D1, 2 = d1, c + d2, c 2
FIGURE 7.33 The tree with three leaves 1, 2, and 3 in addition to an internal node c.
The distances between leaves (D1,2 , D1,3 , and D2,3 ) uniquely define the lengths of
edges (d1,c , d2,c , and d3,c ).
The reason we failed to fit a tree to the 4 ⇥ 4 distance matrix in Figure 7.34 is partly
because the number of equations was larger than the number of variables (di,j ) for each
tree (a fact that will also hold for larger values of n). If the number of equations in a
linear system of equations is smaller than or equal to the number of variables in the
system, then a solution usually exists, whereas if the number of equations exceeds the
number of variables, there is usually no solution. However, there are exceptions in both
directions. Consult an introductory linear algebra text for more details.
The four point condition
The four point condition gives an alternative way to determine if a matrix is additive.
Consider the tree containing only four leaves in Figure 7.35. For this tree, observe that
di,j + dk,l  di,k + d j,l = di,l + d j,k
because the first sum is the sum of lengths of all edges in the tree minus the length of
the internal edge, while the last two sums are equal to the sum of lengths of all edges in
the tree plus the length of the internal edge.
Indeed, for any quartet of leaves (i, j, k, l ) in an arbitrary tree, if we compute the
three sums
di,j + dk,l di,k + d j,l di,l + d j,k
then we will find that two of the sums are equal and that the third sum is less than or
equal to the other two sums. In terms of an n ⇥ n distance matrix, we say that a quartet
of indices (i, j, k, l ) satisfies the four point condition if two of the following sums are
equal, and the third sum is less than or equal to the other two sums:
52
Di,j + Dk,l Di,k + D j,l Di,l + D j,k
Four Point Theorem: A distance matrix is additive if and only if the four point condition
holds for every quartet (i, j, k, l ) of indices of this matrix.
di,a + dj,a = Di,j = 3

i k di,a + da,b + dk,b = Di,k = 4
di,a + da,b + dl,b = Di,l = 3
a b ! dj,a + da,b + dk,b = Dj,k = 4
j l dj,a + da,b + dl,b = Dj,l = 5

dk,b + dl,b = Dk,l = 2
di,a + da,b + dj,b = Di,j = 3

i j di,a + dk,a = Di,k = 4
di,a + da,b + dl,b = Di,l = 3
a b ! dj,b + da,b + dk,a = Dj,k = 4
k dj,b + dl,b = Dj,l = 5

l
dk,a + da,b + dl,b = Dk,l = 2
di,a + da,b + dj,b = Di,j = 3

i j di,a + da,b + dk,b = Di,k = 4
di,a + dl,a = Di,l = 3
a b ! dj,b + dk,b = Dj,k = 4
k dj,b + da,b + dl,b = Dj,l = 5

l
dk,a + da,b + dl,a = Dk,l = 2
j
di,c + dj,c = Di,j = 3
di,c + dk,c = Di,k = 4
di,c + dl,c = Di,l = 3
i c k ! dj,c + dk,c = Dj,k = 4
dj,c + dl,c = Dj,l = 5
dk,c + dl,c = Dk,l = 2
l
FIGURE 7.34 (Left) All four trees with four leaves. (Right) Each attempt to fit the distance
matrix in Figure 7.9 to a tree results in a system of six linear equations. Because none
of these systems has a solution, this matrix must be non-additive.
53
CHAPTER 7
EXERCISE BREAK: Prove the Four Point Theorem.
i k i k
di, k di, l
dj, l dj, k
j l j l
i k
di, j dk, l
j l
FIGURE 7.35 Three pairs of paths through a tree with four leaves. The paths in the top
left and top right traverse the same edges, and so di,k + dj,l = di,l + dj,k . Furthermore,
di,j + dk,l must be less than or equal to these two sums, since it does not traverse the
internal edge of the tree, as shown in the tree on the bottom.
The Four Point Theorem provides us with an alternative way of determining whether
a given distance matrix is additive, since we can simply check whether the four point
condition holds for each quartet of indices of the matrix.
EXERCISE BREAK: Compare the running time of this proposed method with
that of a variant of A DDITIVE P HYLOGENY deciding whether a given distance
matrix is additive.
EXERCISE BREAK: Find a quartet of indices from the distance matrix in Fig-
ure 7.9 (right) that violates the four points condition.
Did bats give us SARS?
During the search for the animal reservoir of the SARS virus, biologists discovered
infected palm civets at a live animal market in China. Meat from these animals is often
added to “dragon-tiger-phoenix soup”, an expensive Cantonese dish. The discovery
did not greatly change the infected civets’ fate: instead of ending up in soup, they were
made into SARS scapegoats and slaughtered anyway.
54
Yet when further searches failed to identify more SARS-infected civets, biologists
started to wonder whether palm civets really were the original source of SARS. In
2005, they discovered a SARS-like virus in Chinese horseshoe bats (Figure 7.36). The
bats turned out to be SARS-CoV carriers, but they can probably only pass the virus
to humans through intermediate hosts. Since bat meat is considered a delicacy and is
also used in traditional Chinese medicine, bats had plenty of chances to come in close
contact with civets at overcrowded live animal markets.
FIGURE 7.36 The horseshoe bat.
When biologists constructed the evolutionary tree of coronaviruses from bats, civets,
and humans (Figure 7.37), they found that both the civet and human variants of SARS-
CoV are nested within a bat virus phylogeny.
Bat 1
Civet 1
Human 1
Human 2
Bat 2
Bat 3
Bat 4
Bat 5
FIGURE 7.37 An evolutionary tree of coronaviruses from bats, civets, and humans.
55
CHAPTER 7
Even before sequencing SARS-CoV, biologists knew of other human coronaviruses

but considered them harmless. However, a new deadly SARS-like coronavirus emerged
in 2012 in Saudi Arabia, This virus (causing Middle East Respiratory Syndrome
(MERS)) generated international headlines as it quickly spread to other countries.
Although researchers initially believed that camels were to blame for MERS — many
Saudis consume unpasteurized camel milk — most afflicted patients had not come in
contact with camels. Yet when researchers tested a coronavirus taken from a bat found
just a few miles from one of the first MERS patient’s homes, they found this virus to be
a nearly perfect match with the patient’s virus sample.
Why does the neighbor-joining algorithm find neighboring leaves?
We stated in the main text that if the distance matrix D is additive, then there exists a
unique simple tree T REE ( D ) fitting this matrix. This is not quite true, since if a simple
tree fits D and has an internal edge of weight zero, then we can easily remove this edge
by “gluing” together the nodes that it connects (Figure 7.38). Thus, we will make a
further assumption that not only is T REE ( D ) simple, but that it contains no internal
edges of length zero.
i k i k
0 7 0 7
0
4 3 4 3
j l j l
FIGURE 7.38 Gluing the nodes at the endpoints of an internal edge having zero length
(shown by the dashed edge).
Now, we can rewrite the neighbor-joining matrix formula as follows:

⇤
Di,j = (n 2) · Di,j T OTAL D ISTANCE D (i ) T OTAL D ISTANCE D ( j)
= (n 2) · Di,j Â Di,k Â D j,k .
1 k  n 1 k  n
Each element in this formula further breaks down as a sum of edge weights in T REE ( D ).
56
For example, for the first tree in Figure 7.34,

⇤
D1,2 = 2 · d1,2 (d1,3 + d1,4 + d1,2 ) (d2,3 + d2,4 + d1,2 )
= 2 · (d1,a + d a,2 ) ((d1,a + d a,b + db,3 ) + (d1,a + d a,b + db,4 )+
(d1,a + d a,2 )) ((d2,a + d a,b + db,3 ) + (d2,a + d a,b + db,4 ) + (d1,a + d a,2 ))
= 2 · d1,a 2 · d2,a 2 · db,3 2 · db,4 4 · d a,b ,
and
⇤
D1,3 = 2 · d1,3 (d1,2 + d1,4 + d1,3 ) (d3,4 + d3,2 + d3,1 )
= 2 · (d1,a + d a,b + db,3 ) ((d1,a + d a,2 ) + (d1,a + d a,b + db,4 )+
(d1,a + d a,b + db,3 )) ((d3,b + db,4 ) + (d3,b + db,a + db,2 ) + (d3,b + db,a + d a,1 ))
= 2 · d1,a 2 · d2,a 2 · db,3 2 · db,4 2 · d a,b .
Note that the expressions for D1,2 ⇤ and D ⇤ are nearly identical, with only the
1,3
⇤
coefficients of d a,b differing (highlighted in red above). Since D1,2 ⇤ = 2·d
D1,3 a,b < 0,
⇤
the neighbor-joining algorithm will prefer the smaller D1,2 over D1,3 .⇤
Given an edge e in T REE ( D ), the multiplicity of this edge in Di,j⇤ is the coefficient
⇤
of de in Di,j , denoted M ULTIPLICITYi,j (e),. For example, for the edge e = ( a, b) in the
first tree in Figure 7.34, M ULTIPLICITY1,2 (e) = 4, and M ULTIPLICITY1,3 (e) = 2. The
following result shows that all limbs in T REE ( D ) have the same multiplicity.
Lemma. For an additive distance matrix D and any pair of leaves i and j in T REE ( D ),
M ULTIPLICITYi,j (e) is equal to 2 for any limb e in T REE ( D ).
Proof. If a limb e is not the limb of leaf i or leaf j, then it is counted zero times in
(n 2) · Di,j , once in T OTAL D ISTANCE D (i ) and once in T OTAL D ISTANCE D ( j), making
its multiplicity 2. On the other hand, if e is the limb of i or j (say, i), then it is
counted n 2 times in (n 2) · Di,j , n 1 times in T OTAL D ISTANCE D (i ), and once in
T OTAL D ISTANCE D ( j). Therefore, its multiplicity is n 2 (n 1) 1 = 2.
This lemma implies that regardless of which pair of leaves i and j we choose, the
limbs of T REE ( D ) will all have the same contribution to the computation of Di,j ⇤ . As a
result, only the multiplicities of internal edges of T REE ( D ) differentiate values of the
matrix D ⇤ . Can we determine these multiplicities?
EXERCISE BREAK: Prove that for any i and j, and for any internal edge e in
T REE ( D ), M ULTIPLICITYi,j (e)  2.
57
CHAPTER 7
Not only does the multiplicity of each internal edge not exceed -2, but we also have
a condition determining when an internal edge will have multiplicity equal to -2. To
derive this condition, first note that the removal of an internal edge e disconnects any
tree into two subtrees. If e lies on the (unique) path connecting leaves i and j in T REE ( D ),
denoted PATH (i, j), then i and j belong to different subtrees, denoted Ti and Tj , respec-
tively; otherwise, i and j belong to the same subtree (Figure 7.39). In the latter case,
we denote the number of leaves in the subtree that does not include leaves i and j as
L EAVESi,j (e).
i e
j
l
FIGURE 7.39 If an internal edge e lies on the unique path connecting two leaves in the
original tree (such as j and k), then these leaves are separated into different subtrees
(shown as blue and green) after removing e. If e does not lie on the unique path
connecting two leaves (such as k and l), then these leaves belong to the same subtree
after removing e.
Edge Multiplicity Theorem: Given an additive matrix D, the multiplicity of an internal

edge e is equal to 2 if e lies on PATH (i, j) in T REE ( D ) and is equal to 2 · L EAVESi,j (e)
otherwise.
Proof. If an internal edge e lies on PATH (i, j) in T REE ( D ), then e has a coefficient of n 2
in (n 2) · Di,j term in Di,j ⇤ . To compute M ULTIPLICITY ( e ), consider the subtrees T
i,j i
and Tj formed by the removal of e. For each leaf k in Ti , PATH ( j, k) passes through e,
thus contributing 1 to the coefficient of e in T OTAL D ISTANCE D ( j), but PATH (i, k) does
not pass through e, thus contributing 0 to the coefficient of e in T OTAL D ISTANCE D (i )).
Likewise, for each leaf k in Tj , PATH (i, k) passes through e, contributing 1 to the coef-
ficient of e in T OTAL D ISTANCE D (i ), but PATH ( j, k) does not pass through e, contributing
0 to the coefficient of e in T OTAL D ISTANCE D (i )). As a result, every leaf k will contribute
1 to the coefficient of e either in T OTAL D ISTANCE D (i ) or in T OTAL D ISTANCE D ( j). The
coefficient of e in T OTAL D ISTANCE D (i ) + T OTAL D ISTANCE D ( j) is n, which means that
58
M ULTIPLICITYi,j (e) = (n 2) n= 2.
On the other hand, if e does not lie on PATH (i, j), then e has a coefficient of 0 in
(n 2) · Di,j . And if k is a leaf in the subtree not containing i and j, then to reach k,
we must pass through e. Thus, the coefficient of e in each of T OTAL D ISTANCE D (i ) and
T OTAL D ISTANCE D ( j) is equal to L EAVESi,j (e). It follows that
M ULTIPLICITY (e) = 0 2 · L EAVESi,j (e) = 2 · L EAVESi,j (e) .
We can interpret the Edge Multiplicity Theorem as stating that internal edges on
the path PATH (i, j) have large multiplicities ( 2) and other internal edges have small
⇤ (among all possible
multiplicities (less than 2). Thus, if we are trying to minimize Di,j
choices of i and j), then we should look for a pair of leaves (i, j) having few internal
edges on PATH (i, j). Neighbors have no internal edges connecting them, which makes
them attractive candidates. The following exercise will get us part of the way toward
proving that neighbors indeed minimize Di,j ⇤.
EXERCISE BREAK: Show that if leaves i and j are neighbors in T REE ( D ), and
⇤ < D⇤ .
leaf k is not a neighbor of i, then Di,j i,k
Neighbor-Joining Theorem: Given an additive distance matrix D, a minimum element Di,j ⇤

⇤
of the neighbor-joining matrix D corresponds to neighboring leaves i and j in T REE ( D ).
Proof. Assume that Di,j ⇤ is a minimum element of D ⇤ but that i and j are not neighbors
in T REE ( D ). We aim to reach a contradiction by finding a pair of neighbors k and l such

⇤ < D ⇤ . By the preceding exercise, neither i nor j can have a neighbor if D ⇤ is a
that Dk,l i,j i,j
minimum element of D ⇤ . Thus, i and j are the only leaves connected to PARENT (i ) and
PARENT ( j), respectively. Because T REE ( D ) is simple, PARENT (i ) and PARENT ( j) have
degree at least equal to 3, meaning that each of these nodes is connected to at least two
other nodes of T REE ( D ), one of which lies on PATH (i, j). The other node is part of its
own subtree; we call these subtrees T1 and T2 (Figure 7.40).
Without loss of generality, we will assume that the number of leaves in T1 does
not exceed the number of leaves in T2 . Because i and j are not in T1 or T2 , T1 must
therefore contain fewer than n/2 leaves, and the rest of T REE ( D ) must contain more
than n/2 leaves (recall that n refers to the total number of leaves in T REE ( D )). Because i
59
CHAPTER 7
T2
PATH(i, j)
l
T1
FIGURE 7.40 Leaves i and j are not neighbors. The unique path connecting them
in TREE(D), PATH (i, j), is shown in blue. Because TREE (D) is simple, PARENT (i) and
PARENT (j) must be connected to at least two other internal nodes, thus forming subtrees
T1 and T2 (shown in red and green).
has no neighbors, T1 must have at least two leaves, which implies that T1 has a pair of
⇤ < D⇤ .
neighbors, which we denote as (k, l ). We will show that Dk,l i,j
⇤  D ⇤ by
Consider an internal edge e of T REE ( D ). We will first show that Dk,l i,j
⇤ does not exceed the multiplicity of e in D ⇤ .
showing that the multiplicity of e in Dk,l i,j
There are three possibilities.
• If e lies on PATH (i, j), then by the Edge Multiplicity Theorem, M ULTIPLICITYi,j (e) =
2, and the result follows.
• If e lies on PATH (i, k ), then the removal of e breaks T REE ( D ) into two subtrees,
one containing k and l (with number of leaves equal to L EAVESi,j (e) < n/2), and
the other containing i and j (with number of leaves equal to L EAVES k,l (e) >
n/2). Thus, by the Edge Multiplicity Theorem, M ULTIPLICITY k,l (e) = 2 ·
L EAVES k,l (e) < M ULTIPLICITYi,j (e) = 2 · L EAVESi,j (e).
60
• If e lies on neither PATH (i, j) nor PATH (i, k), then the subtree not containing i
and j when e is removed is the same as the subtree not containing k and l. As
a result, we have that L EAVESi,j (e) = L EAVES k,l (e), which in turn implies that
M ULTIPLICITYi,j (e) = M ULTIPLICITY k,l (e) (the Edge Multiplicity Theorem).
To prove that Dk,l⇤ is in fact less than D ⇤ , note that PATH (i, k ) must contain an internal
i,j
edge e because i and k are not neighbors. By the middle case above, we know that
M ULTIPLICITY k,l (e) < M ULTIPLICITYi,j (e). Our assumption that no internal edges of
T REE ( D ) have length zero means that de must be positive, which yields the result.
Computing limb lengths in the neighbor-joining algorithm
In the main text, we were attempting to assign limb lengths to leaves of a tree constructed
from an arbitrary distance matrix. We set the limb length of i equal to 12 ( Di,j + Di,j ) and
the limb length of j equal to 12 ( Di,j Di,j ), where
T OTAL D ISTANCE D (i ) T OTAL D ISTANCE D ( j)

Di,j = .
n 2
Where do these formulas come from?
Assume for a moment that D is an additive matrix, and select some leaf k that is not
equal to i or j. If m is the parent of i and j, then we already have the limb length formula
Di,j + Di,k D j,k
L IMB L ENGTH (i ) = .
2
As a result, it might seem like we should use this formula in the neighbor-joining
algorithm for an arbitrary distance matrix D. However, if D is non-additive, then the
expression ( Di,j + Di,k D j,k )/2 will vary depending on how we select k. So we need a
formula that still computes L IMB L ENGTH (i ) when D is additive but that also gives us a
single value when D is non-additive. To this end, we can compute the average of the
above formula over all choices of n 2 leaves k:
1 Di,j + Di,k D j,k
n 2
· Â 2
all leaves k 6= i, j
If D is additive, then the sum above contains n 2 terms, all of which are equal to
L IMB L ENGTH (i ). Furthermore, if D is non-additive, then this formula provides us with
an estimate of the limb length. Note also that the sum above has n 2 occurrences of
61
CHAPTER 7
Di,j . If we separate these out, then we obtain
Di,j 1 Di,k D j,k

L IMB L ENGTH (i ) =
2
+ · Â
n 2 all leaves k 6= i, j 2
!
Di,j 1 Di,k D j,k
=
2
+
n 2
· Â 2 Â 2
all leaves k 6= i, j all leaves k 6= i, j
!!
1 1
= ·
2
Di,j +
n 2
· Â Di,k Â D j,k
all leaves k 6= i, j all leaves k 6= j
✓ ◆
1 T OTAL D ISTANCE D (i ) T OTAL D ISTANCE D ( j)
= · Di,j +
2 n 2
1
= · Di,j + Di,j ,
2
which is the formula in the main text that we used for computing L IMB L ENGTH (i).
Giant panda: bear or raccoon?
For many years, biologists could not agree on whether the giant panda should be
classified as a bear or as a raccoon. Although giant pandas look like bears, they have
features that are unusual for bears and typical of raccoons: they do not hibernate in the
winter, and their male genitalia are tiny and backward-pointing. As a result, Edwin
Colbert wrote in 1938:
So the quest has stood for many years with the bear proponents and the raccoon
adherents and the middle-of-the-road group advancing their several arguments with
the clearest of logic, while in the meantime the giant panda lives serenely in the
mountains of Szechuan with never a thought about the zoological controversies he
is causing by just being himself.
Whereas analysis of anatomical and behavioral characters only led to unsettled

debates, the analysis of genetic characters by Stephen O’Brien in 1985 demonstrated
that giant pandas are indeed more closely related to bears than raccoons (Figure 7.41).
Where did humans come from?
In 1987, Rebecca Cann, Mark Stoneking and Allan Wilson constructed an evolutionary
tree of mitochondrial DNA (mtDNA) of 133 people representing African, Asian, Aus-
tralian, Caucasian, and New Guinean ethnic groups. This tree led to the Out of Africa
62
40
35
30
25
Millions 20
of years ago
15
10
0
Brown Polar Black Spectacled Giant Red
Raccoon
Bear Bear Bear Bear Panda Panda
FIGURE 7.41 An evolutionary tree of bears and raccoons.
hypothesis, which claims that humans have a common ancestor who lived in Africa.
This study turned the question of human origins into an algorithmic problem.
The mtDNA evolutionary tree showed a trunk splitting into two major branches
(Figure 7.42). One branch, containing the bottom five individuals in Figure 7.42, con-
sisted only of Africans, whereas the other branch included some modern Africans as
well as all people belonging to other ethnic groups.
STOP and Think: After looking at the tree in Figure 7.42, where do you think all
humans came from?
If humans populated Africa before Asia, then African genomes started to diverge from
each other earlier than Asian genomes. Thus, we would expect to find that African
genomes, which had more time to diverge from each other, have more mutations
(compared to each other and other genomes) than Asian genomes. This reasoning gives
us a hint on how to check whether the human race spread from Africa.
63
CHAPTER 7
Chukchi
Australian
Australian
Piman
Italian
Papua New Guinea Highland
Papua New Guinea Coast
Papua New Guinea Highland
Georgian
German
Uzbek
Saam
Crimean Tatar
Dutch
French
English
Samoan
Korean
Chinese
Asian Indian
Chinese
Papua New Guinea Coast
Australian
Evenki
Buriat
Kyrgyz
Warao
Warao
Siberian Inuit
Guarani
Japanese
Non-African Japanese
Mkamba
African Ewondo
Bamileke
Lisongo
Yoruba
Yoruba
Mandenka
Ibo
Ibo
Mbenzele
Biaka
Biaka
Mbenzele
Kikuyu
Hausa
Mbuti
Mbuti
San
San
FIGURE 7.42 Evolutionary tree constructed for various human mitochondrial genomes
from Africa (red), Asia (blue), North and South America (purple), Europe (green), and
Oceania (orange). The dashed line divides African from non-African genomes.
African genomes are indeed more diverse than genomes from other continents,
which led Wilson and colleagues to conclude that the African lineage is the oldest and
that modern humans trace their roots back to Africa. Thus, a population of Africans, the
first modern humans, forms one subtree, whereas another subtree represents a subgroup
that left Africa and later spread out to the rest of the world. Using the mitochondrial
tree, Wilson and colleagues further estimated that humans emerged from Africa 130,000
64
years ago, with racial differences arising only 50,000 years ago. Figure 7.43 shows
putative human migration patterns derived from genomic data.
FIGURE 7.43 Putative routes of human migration derived from genomic data labeled
by the number of years ago these migrations occurred.
65
CHAPTER 7
Bibliography Notes
Zuckerkandl and Pauling, 1965 published “Molecules as documents of evolutionary

history”. The evolutionary tree that resolved the giant panda riddle was constructed
by O’Brien et al., 1985. The Out of Africa hypothesis was proposed by Cann, Stoneking,
and Wilson, 1987. The studies of the origins of HIV and the primate reservoir of the HIV
virus were initiated by Gao et al., 1999. The first molecular evidence of HIV transmission
in a criminal case was presented by Metzker et al., 2002. Whiting, Bradler, and Maxwell,
2003 published the study of loss and recovery of wings in stick insects.
The UPGMA approach to evolutionary tree reconstruction was developed by Sokal
and Michener, 1958. The neighbor joining algorithm was developed by Saitou and Nei,
1987. Studier and Keppler, 1988 proved that this algorithm solves the Distance-Based
Phylogeny Problem for additive trees. The dynamic programming algorithm solving
the Small Parsimony problem was developed by Sankoff, 1975. The four point condition
was formulated by Zaretskii, 1965. The nearest neighbor interchange approach to
exploring trees was proposed by Robinson, 1971. Felsenstein, 2004 provides excellent
coverage of various algorithms for tree reconstruction.
66
67
HOW DID YEAST BECOME A WINE MAKER?
An Evolutionary History of Wine Making
How long have we been addicted to alcohol?
One of the first organisms that humans domesticated was yeast. In 2011, while exca-
vating an old graveyard in an Armenian cave, scientists discovered a 6,000 year-old
winery, complete with a wine press, fermentation vessels, and even drinking cups.
This winery was a major technological innovation that required understanding how to
control Saccharomyces, the genus of yeast used in alcohol and bread production.
Yet our interest in alcohol may date back much farther than 6,000 years. In 2008,
scientists discovered that pen-tailed tree shrews (Figure 8.1), which are similar to the
ancient ancestors of all primates, are alcoholics. Their beverage of choice? A “palm wine”
produced by the flowers of Bertram palms and naturally fermented by Saccharomyces
yeast living on the flowers. This finding suggests that our own taste for alcohol may
have a genetic basis that predates the Armenian wine cave by millions of years!
FIGURE 8.1 The pen-tailed tree shrew.
Pound for pound, the amount of palm wine that tree shrews consume would be
lethal to most mammals. Fortunately, the shrews have developed efficient ways of
metabolizing alcohol, and so they avoid inebriation, which would increase their risk
of being killed by a predator. Because of tree shrews’ alcohol tolerance, scientists
believe that alcohol probably offers the shrews some evolutionary advantages, such as
protection against heart attack. It is also possible that our more recent primate ancestors
were also heavy drinkers — after all, chimpanzees drink naturally brewed fruit nectar
— and that we may have inherited an ancestral association of alcohol intake with caloric
gain.
69
CHAPTER 8
The diauxic shift
The species of yeast that we will consider in this chapter is Saccharomyces cerevisiae,
which can brew wine because it converts the glucose found in fruit into ethanol. We
will therefore begin with a simple question: if S. cerevisiae often lives on grapevines,
why must crushed grapes be stored in tightly sealed barrels in order to make wine?
Once its supply of glucose runs out, S. cerevisiae must do something to survive. It
therefore inverts its metabolism, with the ethanol that it just produced becoming its
new food supply. This metabolic inversion, called the diauxic shift, can only occur in
the presence of oxygen. Without oxygen, S. cerevisiae hibernates until either glucose
or oxygen becomes available. In other words, if winemakers don’t seal their barrels,
then the yeast in the barrel will metabolize the ethanol that it just produced, ruining the
wine.
The diauxic shift is a complex process that affects the expression of many genes.
Accordingly, it must derive from a major evolutionary event that equipped the Sac-
charomyces ancestor with a formidable advantage over its competitors: not only can
Saccharomyces kill its competitors by producing ethanol, which is toxic to most bacteria
and other yeasts, but it can then use the accumulated ethanol as an energy source. But
how and when did Saccharomyces invent the diauxic shift? And which genes does it
involve?
Identifying Genes Responsible for the Diauxic Shift
Two evolutionary hypotheses with different fates
Remember Susumu Ohno and his Random Breakage Model? Ohno also hypothesized
that there exist rare evolutionary events called whole genome duplications, or WGDs,
which duplicate an entire genome. He proposed this WGD Model in 1970, when
there was absolutely no evidence to support it. Yet he believed that a WGD would be
needed at the time of a critical evolutionary innovation for a species to implement some
revolutionary new function, such as the diauxic shift.
For example, imagine the time, millions of years ago, when the first fruit-bearing
plants had evolved, but no organisms could metabolize the glucose produced by these
fruits and take advantage of the produced ethanol. The first species to do so would
have possessed an enormous evolutionary advantage, but metabolizing glucose — let
alone ethanol — is not a simple task. Rather than creating a new gene here or there,
the diauxic shift would have required creating new metabolic pathways with many
70
genes working together. Ohno argued that a WGD would provide a platform for such
a revolutionary innovation, since every duplicated gene would have two copies. One
copy would be free to evolve without compromising the gene’s existing function, which
would be carried out by the remaining copy.
The Random Breakage Model and the WGD Model had very different fates. From
the time of its proposal, the Random Breakage Model was embraced by biologists, and
it became dogma until its refutation in 2003. In contrast, the WGD Model was initially
met with skepticism (because only 13% of S. cerevisiae genes are duplicated) and did not
gain traction for 25 years.
In 1997, Wolfe and Shields provided the first computational arguments in favor of
a WGD in S. cerevisiae. They argued that the fact that only 13% of S. cerevisiae genes
are duplicated is not surprising because even if hundreds of genes contribute to an
evolutionary innovation after a WGD, most genes are not needed for this innovation.
Thus, unneeded duplicate genes will be bombarded by mutations until they turn into
pseudogenes and eventually disappear from the genome after millions of years. See
DETOUR: Whole Genome Duplication or a Series of Single-Gene Duplications? to PAGE 111
learn more about the arguments surrounding a WGD in S. cerevisiae.
Which yeast genes drive the diauxic shift?
One of the many steps that yeast performs during fermentation is the conversion
of acetaldehyde into ethanol. If oxygen becomes subsequently available, then the
accumulated ethanol is converted back into acetaldehyde. Both the acetaldehyde-to-
ethanol and ethanol-to-acetaldehyde conversions are catalyzed by an enzyme called
alcohol dehydrogenase (Adh). In S. cerevisiae, Adh activity is encoded by two genes,
Adh1 and Adh2 , which arose from the duplication of a single ancestral gene. The
enzyme encoded by Adh1 has an elevated ability of producing ethanol, whereas the
enzyme encoded by Adh2 has an elevated ability of consuming ethanol.
In 2005, Michael Thomson used a multiple alignment of Adh genes from various
yeast species to reconstruct an ancient Saccharomyces gene. This gene showed a pref-
erence to convert acetaldehyde to ethanol, which resembled the behavior of Adh1 .
Thomson therefore concluded that before the WGD in Saccharomyces, alcohol dehydro-
genase was mainly involved in the generation, not consumption, of ethanol. After the
WGD, Adh1 carried out its original function, whereas Adh2 was free to help power the
diauxic shift.
STOP and Think: How would you find the rest of the genes in S. cerevisae that
work together to accomplish the diauxic shift?
71
CHAPTER 8
Imagine that you were able to monitor all n yeast genes at m time checkpoints on either
side of the diauxic shift, resulting in an n ⇥ m gene expression matrix E, where Ei,j is a
number representing the expression level of gene i at checkpoint j. The i-th row of E
is called the expression vector of gene i. Just by looking at the yeast genes’ expression
vectors, you would observe different patterns of gene behavior with respect to the
diauxic shift. You would see genes whose expression hardly changes, genes whose
expression rapidly increases before the diauxic shift and decreases afterwards, genes
whose expression suddenly increases after the diauxic shift, and so on.
Although this chapter focuses on gene expression with respect to the diauxic shift,
expression matrices are commonplace in biological analysis. For example, if the expres-
sion vector of a newly sequenced gene is similar to the expression vector of a gene with
known function, a biologist may suspect that these genes perform related functions.
Also, genes with similar expression vectors may imply that the genes are co-regulated,
meaning that their expression is controlled by the same transcription factor. This sug-
gests a “guilt by association” strategy for inferring gene functions by starting from a
few genes with known functions and potentially propagating the functions of these
genes to other genes with similar expression vectors.
Finally, gene expression analysis is important in biomedical studies such as ana-
lyzing tissues before and after a drug is administered or contrasting cancerous and
non-cancerous cells. For example, expression analysis led to MammaPrint, a diagnostic
test that determines the likelihood of breast cancer recurrence based on the expression
analysis of 70 human genes associated with tumor activation and suppression.
Yet across all these applications, the question remains: what methods do biologists
use to analyze gene expression data?
Introduction to Clustering
Gene expression analysis
In 1997, Joseph DeRisi conducted the first massive gene expression experiment by sam-
pling an S. cerevisiae culture seven times at hours -6, -4, -2, 0, +2, +4, and +6, where hour
0 indicates the diauxic shift. Since there are approximately 6,400 genes in S. cerevisiae,
this experiment resulted in a 6,400 ⇥ 7 gene expression matrix.
STOP and Think: What technology would you use to generate this matrix?
72
We have already encountered three technologies that could be used to generate this
matrix (see DETOUR: Measuring Gene Expression), but none of these technologies PAGE 111
had matured by 1997! For this reason, DeRisi had to use microarrays, which differ from
the DNA arrays that we discussed in Chapter 2 (see DETOUR: Microarrays). Microar- PAGE 112
rays are rarely used today, but the algorithmic approaches deRisi used for microarray
analysis work equally well for modern gene expression technologies.
STOP and Think: Figure 8.2 visualizes the expression vectors of three yeast
genes. Which of these genes do you think are involved in the diauxic shift?
16
#(&" 16
#(&" 16
#(&"
8
!#%" 8
!#%" 8
!#%"
4
&$" 4
&$" 4
&$"
2
'#" 2
'#" 2
'#"
1
!&" 1
!&" 1
!&"
1/2
%" 1/2
%" 1/2
%"
1/4
$" 1/4
$" 1/4
$"
1/8
#" 1/8
#" 1/8
#"
1/16
!" 1/16
!" 1/16
!"
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
Hours Hours Hours
4
#(&" 4
#(&" 4
#(&"
3
!#%" 3
!#%" 3
!#%"
2
&$" 2
&$" 2
&$"
1
'#" 1
'#" 1
'#"
0
!&" 0
!&" 0
!&"
-1%" -1%" -1%"
-2$" -2$" -2$"
-3#" -3#" -3#"
-4!" -4!" -4!"
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
Hours Hours Hours
FIGURE 8.2 (Top) Expression vectors (1.07, 1.35, 1.37, 3.70, 4.00, 10.00, 5.88),
(1.06, 0.83, 0.90, 0.44, 0.33, 0.13, 0.12), and (1.11, 1.11, 1.12, 1.06, 1.05, 1.06, 1.05)
of three yeast genes (YLR258W, YPL012W, and YPR055W, respectively) visualized
as plots. Each expression vector (e1 , . . . , em ) is represented as a collection of line
segments connecting points (j, ej ) to (j + 1, ej+1 ) for each j between 1 and m 1 = 6.
In the DeRisi experiment, the expression level at the initial checkpoint corresponds
to the base level of expression; note that it is close to 1 for the three genes. Values
above 1 in expression vectors correspond to increased expression, while values below
1 correspond to decreased expression. (Bottom) The expression vectors of the same
three genes with expression levels substituted by their base-2 logarithms: (0.11, 0.43,
0.45, 1.89, 2.00, 3.32, 2.56), (0.09, -0.28, -0.15, -1.18, -1.59, -2.96, -3.08), and (0.15,
0.15, 0.17, 0.09, 0.07, 0.09, 0.07).
73
CHAPTER 8
Note that the pattern of the expression vector of gene YPR055W (Figure 8.2 (top
left)) remains flat during the diauxic shift. We therefore conclude that this gene is
probably not involved in the diauxic shift. On the other hand, the expression of gene
YLR258W (Figure 8.2 (top right)) significantly changes during the diauxic shift, leading
us to hypothesize that this gene is involved in the diauxic shift. Indeed, checking
the Saccharomyces Genome Database (http://yeastgenome.org) reveals that YLR258W
is glycogen synthase. This enzyme controls the production of glycogen, a glucose
polysaccharide that is the main storage vessel for glucose in yeast cells.
In practice, biologists often take the logarithm of expression values (Figure 8.2
(bottom)). After this transformation, positive values of a gene’s expression vector corre-
spond to increased expression, and negative values correspond to decreased expression.
Figure 8.3 shows an expression matrix of ten yeast genes after taking logarithms.
Gene Expression Vector

YLR361C 0.14 0.03 -0.06 0.07 -0.01 -0.06 -0.01
YMR290C 0.12 -0.23 -0.24 -1.16 -1.40 -2.67 -3.00
YNR065C -0.10 -0.14 -0.03 -0.06 -0.07 -0.14 -0.04
YGR043C -0.43 -0.73 -0.06 -0.11 -0.16 3.47 2.64
YLR258W 0.11 0.43 0.45 1.89 2.00 3.32 2.56
YPL012W 0.09 -0.28 -0.15 -1.18 -1.59 -2.96 -3.08
YNL141W -0.16 -0.04 -0.07 -1.26 -1.20 -2.82 -3.13
YJL028W -0.28 -0.23 -0.19 -0.19 -0.32 -0.18 -0.18
YKL026C -0.19 -0.15 0.03 0.27 0.54 3.64 2.74
YPR055W 0.15 0.15 0.17 0.09 0.07 0.09 0.07
FIGURE 8.3 A 10 ⇥ 7 submatrix of DeRisi’s 6,400 ⇥ 7 gene expression matrix for ten
yeast genes (after taking the base-2 logarithm of each expression value). The genes
from Figure 8.2 are colored appropriately.
Clustering yeast genes
Our goal is to partition the set of all yeast genes into k disjoint clusters so that genes
in the same cluster have similar expression vectors. In practice, the number of clusters
is not known a priori, and so biologists typically apply clustering algorithms to gene
expression data for various values of k, selecting the value of k that makes sense
biologically. For simplicity, we will assume that k is fixed. Figure 8.4 shows a partition
of the genes from Figure 8.3 into three clusters indicating increased, decreased, and flat
expression during the diauxic shift.
74
Although the diauxic shift is an important event in the life of S. cerevisiae, it has no
bearing on most of the yeast’s functions. We therefore suspect that most S. cerevisiae
genes have flat expression during the diauxic shift, and we would like to exclude these
genes from further consideration, thus reducing the size of the expression matrix.
Gene Expression Vector

YLR361C 0.14 0.03 -0.06 0.07 -0.01 -0.06 -0.01
YMR290C 0.12 -0.23 -0.24 -1.16 -1.40 -2.67 -3.00
YNR065C -0.10 -0.14 -0.03 -0.06 -0.07 -0.14 -0.04
YGR043C -0.43 -0.73 -0.06 -0.11 -0.16 3.47 2.64
YLR258W 0.11 0.43 0.45 1.89 2.00 3.32 2.56
YPL012W 0.09 -0.28 -0.15 -1.18 -1.59 -2.96 -3.08
YNL141W -0.16 -0.04 -0.07 -1.26 -1.20 -2.82 -3.13
YJL028W -0.28 -0.23 -0.19 -0.19 -0.32 -0.18 -0.18
YKL026C -0.19 -0.15 0.03 0.27 0.54 3.64 2.74
YPR055W 0.15 0.15 0.17 0.09 0.07 0.09 0.07
4
"#
3
$#
2
%#
1
&#
0
'#
-1
!&#
-2
!%#
-3
!$#
-4
!"#
-6 -4 -2 0 2 4 6
Hours
FIGURE 8.4 The rows of the gene expression matrix from Figure 8.3 partitioned into
three clusters. Green genes exhibit increased expression, red genes exhibit decreased
expression, and blue genes exhibit flat behavior and are unlikely to be associated with
the diauxic shift. The element with the largest absolute value in each expression vector
is shown in bold. (Bottom) The rows of the matrix visualized as plots.
The expression levels of most yeast genes hardly change before and after the diauxic
shift (the blue genes in Figure 8.4). These genes also have the property that all of their
expression vector values are very close to zero. In our analysis, we will exclude genes
with expression vectors whose values are all between -2.3 and 2.3. Doing so reduces the
75
CHAPTER 8
original 6,400 yeast gene dataset to a dataset containing 230 genes whose expression
changes significantly around the diauxic shift.
You can see in Figure 8.4 that gene YLR258W has a different pattern of change than
gene YGR043C, indicating that dividing the 230 yeast genes into just two clusters (i.e.,
those with increasing and decreasing expression levels) may be too simplistic. Our goal
is to cluster these genes based on similar patterns of behavior.
The Good Clustering Principle
To identify groups of genes with similar expression patterns, we will think of an

expression vector of length m as a point in m-dimensional space; genes with similar
expression vectors will therefore form clusters of nearby points. Ideally, clusters should
satisfy the following common-sense principle, which is illustrated in Figure 8.5.
Good Clustering Principle: Every pair of points from the same cluster should be closer to
each other than any pair of points from different clusters.
We have therefore embedded gene expression analysis within the algorithmic problem
of partitioning a collection of n points in m-dimensional space into k clusters, which
will be our focus in this chapter.
Good Clustering Problem:

Partition a set of points into clusters.
Input: A set of n points in m-dimensional space and an integer k.

Output: A partition of the n points into k clusters satisfying the Good Clus-
tering Principle.
EXERCISE BREAK: Form ten points in two-dimensional space by taking the

fourth and seventh columns of the matrix in Figure 8.4. How should these points
be partitioned into three clusters?
EXERCISE BREAK: Compute the number of partitions of n points into two

non-empty clusters.
76
FIGURE 8.5 (Left) A partition of twenty points into three clusters that do not satisfy the
Good Clustering Principle. (Right) A different partition of these points that does satisfy
the Good Clustering Principle.
The eye naturally divides the points in Figure 8.6 (left) into two clusters. Unfor-
tunately, these clusters do not satisfy the Good Clustering Principle; in fact, no such
partition of these points into two clusters exists! As a result, we will need to take a dif-
ferent approach in order to devise a well-defined computational problem for clustering.
EXERCISE BREAK: Design a polynomial algorithm to check whether there is a

solution of the Good Clustering Problem.
(8, 7)
(1, 6) (5, 6)
(3, 4)
(1, 3) (10, 3)
(5, 2)
(7, 1)
FIGURE 8.6 (Left) A collection of points that obviously form two clusters but for which
no partition into two clusters satisfies the Good Clustering Principle. (Right) Eight points
in two-dimensional space.
77
CHAPTER 8
STOP and Think: Figure 8.6 (right) shows eight data points in two-dimensional
space. How would you partition these points into three clusters? How can
we transform the Good Clustering Problem into a well-defined computational
problem?
Clustering as an Optimization Problem
Rather than thinking about clustering as dividing data points Data into k clusters, we
will instead try to select a set Centers of k points that will serve as the centers of these
clusters. We would like to choose Centers so that they minimize some distance function
between Centers and Data over all possible choices of centers. But how should this
distance function be defined?
First, we define the Euclidean distance between points v = (v1 , . . . , vm ) and w =
(w1 , . . . , wm ) in m-dimensional space, denoted d(v, w), as the length of the line segment
connecting these points,
q
d(v, w) = Âim=1 (vi wi )2 .
Next, given a point DataPoint in multi-dimensional space and a set of k points Centers,
we define the distance from DataPoint to Centers, denoted d(DataPoint, Centers), as the
Euclidean distance from DataPoint to its closest center,
d(DataPoint, Centers) = min d(DataPoint, x) .

all points x from Centers
The length of the segments in Figure 8.7 correspond to d(DataPoint, Centers) for each
point DataPoint.
We now define the distance between all data points Data and centers Centers. This dis-
tance, denoted M AX D ISTANCE (Data, Centers), is the maximum of d(DataPoint, Centers)
among all data points DataPoint,
M AX D ISTANCE (Data, Centers) = max d(DataPoint, Centers) .

all points DataPoint from Data
In Figure 8.7, this distance corresponds to the length of the red segment. We can now
formulate a well-defined clustering problem.
78
FIGURE 8.7 The collection of points Data from Figure 8.5 (shown as black points)
along with three centers forming a set Centers (shown as colored points). For each
point DataPoint in Data, d(DataPoint, Centers) is equal to the length of the segment
connecting it to its nearest center. MAXDISTANCE (Data, Centers) is equal to the length
of the longest such segment, which is shown in red.
k-Center Clustering Problem:

Given a set of data points, find k centers minimizing the maximum distance between
these data points and centers.
Input: A set of points Data and an integer k.

Output: A set Centers of k centers that minimize the distance
M AX D ISTANCE (DataPoints, Centers) over all possible choices of k centers.
EXERCISE BREAK: How would you select a center in the case of only a single
cluster (i.e., when k = 1)?
Farthest First Traversal
Although the k-Center Clustering Problem is easy to state, it is NP-Hard. The Farthest
First Traversal heuristic, whose pseudocode is shown below, selects centers from the
points in Data (instead of from all m-dimensional points). It first selects an arbitrary
point in Data as the first center and iteratively adds a new center as the point in Data
that is farthest from the centers chosen so far, with ties broken arbitrarily (Figure 8.8).
79
CHAPTER 8
FIGURE 8.8 Applying FARTHESTFIRSTTRAVERSAL to the data in Figure 8.5. (Top left) An
arbitrary point from the dataset (shown in blue) is selected as the first center. All points
belong to a single cluster. (Top right) The red point is selected as the second center,
since it is the farthest from the blue point. (Bottom left) After computing each data
point’s minimum distance to each of the first two centers, we find that the point with
the largest such distance is the green point, which becomes the third center. (Bottom
right) The fourth center is shown in purple.
FARTHESTFIRSTTRAVERSAL(Data, k)
Centers the set consisting of a single randomly chosen point from Data
while |Centers| < k
8A DataPoint the point in Data maximizing d(DataPoint, Centers)
add DataPoint to Centers
return Centers
80
EXERCISE BREAK: Apply FARTHEST F IRST T RAVERSAL to the eight points in

Figure 8.6 (right) with k = 3. How does the result change if you change the point
that is selected first?
EXERCISE BREAK: Let Centers be the set of centers returned by

FARTHEST F IRST T RAVERSAL, and let Centersopt be a set of centers corresponding
to an optimal solution of the k-Center Clustering Problem. Prove that
M AX D ISTANCE (Data, Centers)  2 · M AX D ISTANCE (Data, Centersopt ).
Can you find a collection of data points such that the centers returned by
FARTHEST F IRST T RAVERSAL are suboptimal?
FARTHEST F IRST T RAVERSAL is fast, and according to the preceding exercise, its solu-
tion approximates the optimal solution of the k-Center Clustering Problem; however,
this algorithm is rarely used for gene expression analysis. In k-Center Clustering, we
selected Centers so that these points would minimize M AX D ISTANCE (Data, Centers),
the maximum distance between any point in Data and its nearest center. But biologists
are usually interested in analyzing typical rather than maximum deviations, since the
latter may correspond to outliers representing experimental errors (Figure 8.9).
FIGURE 8.9 (Left) A set of data points with three clearly seen clusters and two outliers.
(Right) Because FARTHESTFIRSTTRAVERSAL relies on MAXDISTANCE to compute new
centers, if we attempt to cluster the data points into three clusters, then regardless of
which point is selected as the first center, the two outliers on the left will be selected as
centers of single-element clusters, and all remaining data points will be assigned to a
single cluster.
81
CHAPTER 8
STOP and Think: Can you devise an alternative scoring function that is more
biologically appropriate than M AX D ISTANCE (Data, Centers)?
k-Means Clustering
Squared error distortion
To address limitations of M AX D ISTANCE, we will introduce a new scoring function.

Given a set Data of n data points and a set Centers of k centers, the squared error
distortion of Data and Centers, denoted D ISTORTION (Data, Centers), is defined as the
mean squared distance from each data point to its nearest center,
1
D ISTORTION (Data, Centers) = Â d(DataPoint, Centers)2 .
n all points DataPoint in Data
Note that whereas M AX D ISTANCE (Data, Centers) only accounts for the length of the
single red segment in Figure 8.7, the squared error distortion accounts for the length of
all segments in this figure.
EXERCISE BREAK: Compute the values of M AX D ISTANCE (Data, Centers) and

D ISTORTION (Data, Centers) for the eight data points and the three centers shown
in Figure 8.10 (left). How do these values differ for the centers in Figure 8.10
(right)?
Squared Error Distortion Problem:

Compute the squared error distortion of a set of data points with respect to a set of centers.
8B
Input: A set of points Data and a set of centers Centers.
Output: The squared error distortion D ISTORTION (Data, Centers).
The squared error distortion leads us to the following modification of the k-Centers
Clustering Problem.
82
k-Means Clustering Problem:

Given a set of data points, find k center points minimizing the squared error distortion.
Input: A set of points Data and an integer k.

Output: A set Centers of k centers that minimize D ISTORTION (Data, Centers)
over all possible choices of k centers.
Although the k-Centers and k-Means Clustering Problems look similar, they may
produce different results (Figure 8.10). The key difference between the k-Centers and
k-Means Clustering Problems is that in the latter, the placement of a center is far less
affected by outliers (Figure 8.11).
(6.5, 6.5)
(9, 5)
(3, 9/2)
(5/3, 13/3)
(22/3, 2)
(6, 3/2)
FIGURE 8.10 (Left) The three colored centers solving the k-Centers Clustering Problem
for the eight black data points from Figure 8.6 (right), along with the clusters that they
form. (Right) The three colored centers solving the k-Means Clustering Problem for
these points, along with the clusters that they form.
k-means clustering and the center of gravity
It turns out that the k-Means Clustering Problem is NP-Hard when k > 1. However,
when k = 1, the k-Means Clustering Problem amounts to finding a single center point x
that minimizes the squared error distortion. Although we acknowledge that partitioning
a set of data points into a single cluster is trivial, it remains unclear how to find a single
center minimizing the squared error distortion. We would like to solve this simpler
problem because it will help us design a heuristic for the case when k > 1.
83
CHAPTER 8
FIGURE 8.11 Center placement varies in different clustering problem formulations.

(Left) In the k-Center Clustering Problem, a cluster’s center is chosen so that the
maximum distance between the center and any point in the cluster is minimized.
As a result, the position of the center can be greatly influenced by outliers. (Right) In the
k-Means Clustering Problem, the outlier’s influence over the placement of the center is
much smaller. This behavior is preferable when analyzing biological datasets, in which
outliers often correspond to erroneous data.
We define the center of gravity of Data as the point whose i-th coordinate is the
average of the i-th coordinates of all points from Data. For example, the center of gravity
of the points (3, 8), (8, 0), and (7, 4) is
✓ ◆
3+8+7 8+3+4
, = (6, 5).
3 3
Center of Gravity Theorem: The center of gravity of a set of points Data is the unique point
solving the k-Means Clustering Problem for k = 1.
PAGE 113 For a proof of this theorem, see DETOUR: Proof of the Center of Gravity Theorem.
EXERCISE BREAK: Although the k-Means Clustering Problem is NP-hard for

k > 1, it can be solved in polynomial time for any value of k in the case of
clustering in one-dimensional space, i.e., when all data points fall on a line.
Design an algorithm for solving the k-Means Clustering Problem in this case.
EXERCISE BREAK: Prove that the centers in Figure 8.10 (right) solve the k-
Means Clustering Problem for k = 3.
84
The Lloyd Algorithm
From centers to clusters and back again
The Lloyd algorithm is one of the most popular clustering heuristics for the k-Means
Clustering Problem. It first chooses k arbitrary points Centers from Data as centers and
then iteratively performs the following two steps (Figure 8.12):
• Centers to Clusters: After centers have been selected, assign each data point to
the cluster corresponding to its nearest center; ties are broken arbitrarily.
8C
• Clusters to Centers: After data points have been assigned to clusters, assign each
cluster’s center of gravity to be the cluster’s new center.
STOP and Think: Is it possible for the Lloyd algorithm to produce two centers
that coincide (thus resulting in fewer than k clusters)?
In Figure 8.12, the centers appear to be moving less and less between iterations. We say
that the Lloyd algorithm has converged if the centers (and therefore their clusters) stop
changing between iterations.
STOP and Think: Can you find a set of data points for which the Lloyd algorithm
does not converge?
If the Lloyd algorithm has not converged, the squared error distortion must decrease in
any step, according to the following reasoning:
• In a “Centers to Clusters” step, if a data point is assigned to a new center, then this
point must be closer to the new center than its previous center. Thus, the squared
error distortion must decrease.
• In a “Clusters to Center” step, if a center is updated as a cluster’s center of gravity,

then by the Center of Gravity Theorem, the new center is the only point minimizing
the squared error distortion for the points in its cluster. Thus, the squared error
distortion must decrease.
STOP and Think: Does this reasoning imply that the Lloyd algorithm must
converge?
85
CHAPTER 8
From Clusters to Centers From Centers to Clusters
1 2
3 4
5 6
FIGURE 8.12 The Lloyd algorithm in action for k = 3. In the top left panel, we select
three arbitrary data points as centers, shown as differently colored points. In subsequent
panels, we iterate the “Centers to Clusters” step followed by the “Clusters to Centers”
step. In the bottom right panel, the Lloyd algorithm has converged.
86
It is not true that the Lloyd algorithm must converge just because the squared error
distortion decreases at each step. For example, it could be the case that subsequent
decreases in squared error distortion become smaller and smaller, leading to an infinite
process (e.g., if the error distortion decreases by 1/2, then 1/4, then 1/8, and so on).
The following exercise ensures that such a scenario cannot occur.
EXERCISE BREAK: Prove that the number of iterations of the Lloyd algorithm
does not exceed the number of partitions of the data points into k clusters.
EXERCISE BREAK: Pick your favorite parameter k and run the Lloyd algorithm
1,000 times on the 230-gene diauxic shift dataset, each time initialized with a new
set of k randomly chosen centers. Construct a histogram of the squared error
distortions of the resulting 1,000 outcomes. How many times did you have to run
the Lloyd algorithm before finding the run that scored highest among your 1,000
runs?
Initializing the Lloyd algorithm
Figure 8.13 illustrates that things can go horribly wrong if we do not pay attention to
the Lloyd algorithm’s initialization step. In Figure 8.13 (top), we select no centers from
clump 1, two centers from clump 3, and one center from each of clumps 2, 4, and 5. As
shown in Figure 8.13 (bottom), after the first iteration of the Lloyd algorithm, all points
in clumps 1 and 2 will be assigned to the red center, which will move approximately
halfway between clumps 1 and 2. The two centers in clump 3 will divide the points in
that clump into two clusters. And the centers in clumps 4 and 5 will move toward the
middle of these clumps. The Lloyd algorithm will then quickly converge, resulting in
an incorrect clustering.
EXERCISE BREAK: Compute the probability that at least one of the five clumps
in Figure 8.13 will have no centers if five centers are chosen randomly from the
data (like in the Lloyd algorithm).
STOP and Think: How would you change the Lloyd algorithm’s initialization
step to improve the clusters it finds?
87
CHAPTER 8
1 2 3 4 5
FIGURE 8.13 (Top) Five clumps of ten points in two-dimensional space. The Lloyd
algorithm is initialized so that clump 1 contains no centers, clump 3 contains two
centers (blue and purple), and each of the other three clumps contains one center (red,
orange, and green). (Bottom) The Lloyd algorithm has combined the points in clumps
1 and 2 into a single cluster and split clump 3 into two clusters.
k-means++ Initializer
We have thus far not paid much attention to how initial centers are chosen in the Lloyd
algorithm, which selects them randomly. Similarly to FARTHEST F IRST T RAVERSAL,
k- MEANS ++I NITIALIZER picks k centers one at a time, but instead of choosing the
point farthest from those picked so far, it chooses each point at random in such a way
that distant points are more likely to be chosen than nearby points. Specifically, the
probability of selecting a center DataPoint from Data is proportional to the squared
distance of DataPoint from the centers already chosen, i.e., to d(DataPoint, Centers)2 .
For a simple example, say that we have just three data points, and that the squared
distances from these points to the existing centers Centers are equal to 1, 4, and 5. Then
the probability of k- MEANS ++I NITIALIZER selecting each of these points as the next
center is 1/10, 4/10, and 5/10, respectively.
k-MEANS++INITIALIZER(Data, k)
Centers the set consisting of a single randomly chosen point from Data
while |Centers| < k
randomly select DataPoint from Data with probability proportional to
d(DataPoint, Centers)2
add DataPoint to Centers
return Centers
STOP and Think: Although k- MEANS ++I NITIALIZER may also fall into the trap
in Figure 8.13, it does so rarely. Why?
88
EXERCISE BREAK: Power up your implementation of the Lloyd algorithm

using k- MEANS ++I NITIALIZER, and apply it to the 230-gene diauxic shift dataset
for varying values of k.
Clustering Genes Implicated in the Diauxic Shift
Since selecting the most biologically relevant value of k can be challenging, we will
(somewhat arbitrarily) choose to cluster the 230 yeast genes into six clusters (Figure 8.14).
Cluster 1 Cluster 2 Cluster 3

4 4 4
4
4
4
2 2 2
2
2
0 0 0
0
0
0
-2 -2 -2
2
2
2
-4 -4 -4
4
4
4

4 4 4
4
4
2 2 2
2
2
2
0 0 0
0
0
0
-2 -2 -2
2
2
2
-4 -4 -4
4
4
4
FIGURE 8.14 Applying the Lloyd algorithm (with k = 6) to the abridged yeast dataset
containing 230 genes results in six clusters revealing six different types of regulatory
behavior and containing 37, 36, 58, 19, 36, and 44 genes. Expression vectors for all
genes in each of these six clusters are visualized as separate plots.
The plots in Figure 8.14 reveal six patterns of behavior of genes involved in the
diauxic shift and raise questions for further biological studies beyond the focus of this
chapter. For example, what regulatory mechanisms force the genes in the first cluster to
increase their expression? What mechanisms cause the genes in the fourth cluster to
decrease their expression? And how do these changes contribute to the diauxic shift?
89
CHAPTER 8
EXERCISE BREAK: The clustering of the entire 6,400-gene yeast dataset into six
clusters implies a clustering of the abridged set containing 230 genes. How does
this implied clustering compare to the clustering in Figure 8.14?
Limitations of k-Means Clustering
After seeing the Lloyd algorithm in action, it may seem that clustering is easy. If you
think so, consider the following question.
STOP and Think: How would you cluster the points in Figure 8.15?
FIGURE 8.15 Difficult clustering problems for k = 2 (left and middle) and k = 3 (right).
In the case of challenging clustering problems, the Lloyd algorithm sometimes fails to
identify what may seem like obvious clusters (Figure 8.16).
STOP and Think: The Lloyd algorithm assigns each data point to its closest cen-
ter, with ties broken arbitrarily. What are the negative effects of this assignment?
One weakness with our formulation of the k-Means Clustering Problem is that it forces
us to make a “hard” assignment of each point to only one cluster. This strategy makes
little sense for midpoints, or points that are approximately equidistant from two centers.
To deal with midpoints, our goal is to transition away from a rigid assignment of a data
point to a single cluster (Figure 8.17 (left)) and toward a “soft” assignment (Figure 8.17
(right)).
90
FIGURE 8.16 The human eye (top) and the Lloyd algorithm (bottom) often disagree in
the case of elongated clusters (left), clusters with non-globular shapes (middle), and
clusters with widely different data point densities (right).
(0.98, 0.02)
(0.48, 0.52)
(0.01, 0.99)
FIGURE 8.17 (Left) The points from Figure 8.6 (left) partitioned into two clusters by the
Lloyd algorithm; points are colored red or blue depending on their cluster membership.
(Right) We can visualize a soft clustering of the same data into two clusters as assigning
each point a pair of numbers representing the point’s percentage of “blue” and “red”
based on each cluster’s “responsibility” for this point. The colors mix to form a color
from the red-blue spectrum.
91
CHAPTER 8
From Coin Flipping to k-Means Clustering
Flipping coins with unknown biases
To develop an algorithm for soft clustering, we will introduce a seemingly unrelated

analogy. Tom Stoppard’s play Rosencrantz and Guildenstern Are Dead opens with the
title characters flipping a coin over and over, finding that it results in “heads” 157
consecutive times. Rosencrantz and Guildenstern question whether they have become
detached from the laws of probability, but if you witnessed such an event, you would
probably guess that the coin was biased. Say that your friend flips a biased coin n times,
and you would like to estimate the probability q that a single coin flip results in heads.
STOP and Think: For each value of q between 0 and 1, you can compute the
probability of a given sequence of flips. How would you estimate the value of q
maximizing this probability after watching a series of flips where the coin lands
on heads i out of n times?
It seems like the best estimate of q should be the number of occurrences of heads divided
by the total number of coin flips. But how can we prove this? Given a sequence of
n coin flips containing i heads, the probability that a coin with bias q generated this
sequence is f(q ) = q i · (1 q )n i . Since the most likely coin bias is the value of q that
maximizes this probability, we will set the derivative of f(q ) equal to zero,
1 i 1
f 0 (q ) = i · q i · (1 q )n i
q i · (n i ) · (1 q )n
1 i 1
= [ i · (1 q) q · (n i )] · q i · (1 q )n
1 i 1
= (i q · n) · q i · (1 q )n = 0.
Other than q = 0 and q = 1, the only solution of this equation is q = i/n, implying that
the observed proportion of heads provides the best estimate for q.
To make the coin flipping problem a bit more interesting, suppose that your friend
secretly switches between two coins A and B that look identical but have unknown
biases q A and q B . After observing a sequence of coin flips, your goal is to estimate q A
and q B , which we collectively denote by Parameters.
We will simplify the problem by assuming that every n flips, your friend secretly
decides to either keep the same coin or switch coins. Five sequences of n = 10 flips
are shown in Figure 8.18; we will represent the proportion of heads in each of these
sequences as a vector,
Data = (Data1 , Data2 , Data3 , Data4 , Data5 ) = (0.4, 0.9, 0.8, 0.3, 0.7) .
92
Data
H T T T H T T H T H 0.4
H H H H T H H H H H 0.9
H T H H H H H T H H 0.8
H T T T T T H H T T 0.3
T H H H T H H H T H 0.7
FIGURE 8.18 Five sequences of ten coin flips result in Data = (0.4, 0.9, 0.8, 0.3, 0.7).
“H” denotes heads and “T” denotes tails.
If you knew that your friend used coin A in the first and fourth sequences of flips,
then you would estimate q A by computing the proportion of heads in these sequences,
Data1 + Data4 0.4 + 0.3
qA = = = 0.35 .
2 2
You would then estimate q B as the proportion of heads in the remaining three sequences,
Data2 + Data3 + Data5 0.9 + 0.8 + 0.7
qB = = = 0.8 .
3 3
We will represent this choice of coins as a binary vector HiddenVector = (1, 0, 0, 1, 0),
where a 1 in the k-th position denotes that coin A was used to generate the k-th sequence
of flips, and a 0 denotes that coin B was used. This notation allows us to rewrite the
equations for Parameters in terms of Data and HiddenVector:
Âi HiddenVectori · Datai 1 · 0.4 + 0 · 0.9 + 0 · 0.8 + 1 · 0.3 + 0 · 0.7
qA = = = 0.35
Âi HiddenVectori 1+0+0+1+0
Âi (1 HiddenVectori ) · Datai 0 · 0.4 + 1 · 0.9 + 1 · 0.8 + 0 · 0.3 + 1 · 0.7
qB = = = 0.80
Âi (1 HiddenVectori ) 0+1+1+0+1
where i runs over all data points.
The expression Âi HiddenVectori · Datai is the dot product of vectors HiddenVector and
!
Data, written HiddenVector · Data. Define the all ones vector, written 1 , as the vector
consisting of all ones and whose length is equal to that of HiddenVector. This allows us
!
to write Âi HiddenVectori as the dot product HiddenVector · 1 and Âi (1 HiddenVectori )
! !
as the dot product ( 1 HiddenVector) · 1 . As a result, the above equations become
HiddenVector · Data
qA = !
HiddenVector · 1
!
( 1 HiddenVector) · Data
qB = ! !
( 1 HiddenVector) · 1
93
CHAPTER 8
STOP and Think: We just saw that given Data and HiddenVector, we can find
Parameters = (q A , q B ). If you are given Data and Parameters, can you find the most
likely choice of HiddenVector?
If we know Parameters, then deciding on the most likely choice of HiddenVector corre-
sponds to determining whether coin A or coin B was more likely to have generated the
n observed flips in each of the five coin flipping sequences. For example, suppose we
know that Parameters = (q A , q B ) = (0.6, 0.82). If coin A was used to generate the fifth
sequence of flips, then the probability that it generated the outcome in Figure 8.18 is
q 7A (1 q A )3 = 0.67 · 0.43 ⇡ 0.00179 .
If coin B was used to generate the fifth sequence, then the probability that it generated
this outcome is
q B7 (1 q B )3 = 0.827 · 0.183 ⇡ 0.00145 .
Since 0.00179 > 0.00145, we would set HiddenVector5 equal to 1.
EXERCISE BREAK: Determine the rest of the entries in HiddenVector for Parame-
ters = (0.6, 0.82) and the sequences of coin flips in Figure 8.18.
More generally, let Pr(Datai |q ) denote the conditional probability of generating the
outcome Datai given a coin with bias q,
Pr(Datai |q ) = q n·Datai (1 q )n·(1 Datai ) .
If Pr(Datai |q A ) > Pr(Datai |q B ), then coin A is more likely to have generated the i-th
sequence of flips, and we set HiddenVectori equal to 1. If Pr(Datai |q A ) < Pr(Datai |q B ),
then coin B is more likely, and we set HiddenVectori equal to 0. Ties are broken arbitrarily.
In summary, if HiddenVector is known and Parameters is unknown, then we can
reconstruct the most likely Parameters = (q A , q B ):
(Data, HiddenVector, ?) ! Parameters
Likewise, if Parameters is known and HiddenVector is unknown, then we can reconstruct

the most likely HiddenVector:
(Data, ?, Parameters) ! HiddenVector
Our original problem, however, was that both HiddenVector and Parameters are unknown:
(Data, ?, ?) ! ???
94
Where is the computational problem?
You may have noticed that we have not formulated the computational problem that we
are trying to solve. So define the conditional probability of generating a sequence of
coin flips Datai given HiddenVector and Parameters as
(
Pr(Datai |q A ) if HiddenVectori = 1
Pr(Datai |HiddenVector, Parameters) =
Pr(Datai |q B ) if HiddenVectori = 0
Furthermore, define the conditional probability of generating Data given HiddenVector

and Parameters as
n
Pr(Data|HiddenVector, Parameters) = ’ Pr(Datai |HiddenVector, Parameters) .
i =1
Given Data, the computational problem we are trying to solve is to find HiddenVector
and Parameters maximizing Pr(Data|HiddenVector, Parameters).
From coin flipping to the Lloyd algorithm
Identifying HiddenVector and Parameters from Data may appear hopeless, but we have
already learned that starting from a random guess is not necessarily a bad idea. We
will therefore start from an arbitrary choice of Parameters = (q A , q B ) and immediately
reconstruct the most likely HiddenVector:
(Data, ?, Parameters) ! HiddenVector
As soon as we know HiddenVector, we will question the wisdom of our initial choice of
Parameters and re-estimate Parameters’:
(Data, HiddenVector, ?) ! Parameters’
As illustrated in Figure 8.19 for the initial choice of Parameters = (0.6, 0.82), we repeat
these two steps and hope that Parameters and HiddenVector are moving closer to the
values that maximize Pr(Data|HiddenVector, Parameters),
(Data, ?, Parameters) ! (Data, HiddenVector, Parameters)
! (Data, HiddenVector, ? )
! (Data, HiddenVector, Parameters’)
! (Data, ? , Parameters’)
! (Data, HiddenVector’, Parameters’)
! ···
95
CHAPTER 8
A = 0.6 A = 0.82
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(Data, Parameters) HiddenVector
A = 0.6 B = 0.82
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(Data, HiddenVector) Parameters
A = 0.467 B = 0.85
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(Data, Parameters) HiddenVector
A = 0.467 B = 0.85
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(Data, HiddenVector) Parameters
A = 0.35 B = 0.8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
FIGURE 8.19 Starting with Parameters = (0.6, 0.82) results in HiddenVector = (1, 0, 0,
1, 1) for Data = (0.4, 0.9, 0.8, 0.3, 0.7). We then update Parameters as (0.467, 0.85),
which in turn results in HiddenVector = (1, 0, 0, 1, 0). This new vector leads to the
assignment of Parameters as (0.35, 0.8), at which point the process has terminated
because HiddenVector will not change in the next step.
EXERCISE BREAK: Prove that this process terminates, i.e., that HiddenVector
and Parameters eventually stop changing between iterations.
STOP and Think: If HiddenVector consists of all zeroes, then there are no flips
with coin A, and the formula for computing q A is invalid. What would you do to
address this complication?
Return to clustering
Figure 8.19 has revealed that although our coin flipping analogy seems like a different
problem, it is really just a one-dimensional clustering problem in disguise!
96
STOP and Think: What are Data, HiddenVector, and Parameters in the Lloyd
algorithm?
Given n data points in m-dimensional space Data = (Data1 , . . . , Datan ), we represent

their assignment to k clusters as an n-dimensional vector
HiddenVector = (HiddenVector1 , . . . , HiddenVectorn ) ,
where each HiddenVectori can take integer values from 1 to k. We will then represent the
k centers as k points in m-dimensional space, Parameters = (q1 , . . . , qk ).
In k-means clustering, similarly to the coin flipping analogy, we are given Data, but
HiddenVector and Parameters are unknown. The Lloyd algorithm starts from randomly
chosen Parameters, and we can now rewrite its two main steps as follows:
• Centers to Clusters: (Data, ?, Parameters) ! HiddenVector
• Clusters to Centers: (Data, HiddenVector, ?) ! Parameters
The only difference between the coin flipping algorithm and the Lloyd algorithm for
k-means clustering is how they execute the “Centers to Clusters” step. In the former, we
compute HiddenVectori by comparing Pr(Datai |q A ) with Pr(Datai |q B ), whereas in the
latter, we assign a point to the cluster containing the center nearest to that point.
STOP and Think: Consider the following questions regarding coin flipping and
clustering.
• Is it fair to always select coin A if Pr(Datai |q A ) is only slightly larger than

Pr(Datai |q B )?
• Is it fair to always assign a point to a center if this center is only slightly

closer to the point than another center?
Making Soft Decisions in Coin Flipping
Expectation maximization: the E-step
We will now use our coin flipping analogy to motivate a soft version of k-means
clustering. Given Parameters = (q A , q B ), we can make hard decisions for HiddenVector
by comparing Pr(Datai |q A ) with Pr(Datai |q B ). But this does not mean that we are
97
CHAPTER 8
certain which coin was used. If Pr(Datai |q B ) were approximately equal to Pr(Datai |q A ),
then our confidence that coin B was used would be approximately 50%. On the other
hand, if Pr(Datai |q B ) were much larger than Pr(Datai |q A ), then we would be almost
positive that coin B was used. More generally, we can speak of our confidence that a
coin was used as the “responsibility” of this coin for a given sequence of flips. (The
responsibilities should sum to 1.)
In terms of k-means clustering, if a data point is a midpoint between two centers,
then each of these centers should have about the same responsibility for attracting it to
their clusters. As with coin flipping, the responsibilities of all centers for a given data
point should sum to 1.
STOP and Think: Given Parameters = (q A , q B ) = (0.6, 0.82) and the sequence of
coin flips “THHHTHHHTH”, how would you compute responsibilities for coins
A and B?
To answer the preceding question, we have seen that Pr(0.7|q A ) = 0.67 · 0.43 ⇡ 0.00179
and that Pr(0.7|q B ) = 0.827 · 0.183 ⇡ 0.00145. Before, we rigidly concluded that coin A
was more likely. Now, since coin A is more likely to generate seven heads in a sequence
of ten coin flips, we should assign a larger responsibility to coin A than to coin B. One
possible way to assign these responsibilities is given by the formulas
Pr(0.7|q A ) 0.00179
= ⇡ 0.55
Pr(0.7|q A ) + Pr(0.7|q B ) 0.00179 + 0.00145
Pr(0.7|q B ) 0.00145
= ⇡ 0.45
Pr(0.7|q A ) + Pr(0.7|q B ) 0.00179 + 0.00145
As a result, instead of a vector HiddenVector, we now have a 2 ⇥ 5 responsibility

profile HiddenMatrix that can be constructed from Data and Parameters,
(Data, ?, Parameters) ! HiddenMatrix .
We call this transition the E-step (Figure 8.20).
98
Data
0.4 0.9 0.8 0.3 0.7
E-step M-step
A= 0.6 0.97 0.12 0.29 0.99 0.55 A= 0.483
B= 0.82 0.03 0.88! 0.71 0.01 0.45 B= 0.813
Parameters HiddenMatrix Parameters’
FIGURE 8.20 In the E-step, we compute HiddenMatrix from Data = (0.4, 0.9, 0.8, 0.3,
0.7) and Parameters = (0.6, 0.82). In the M-step, we compute an updated Parameters0
from HiddenMatrix and Data.
Expectation maximization: the M-step
When making hard assignments, we computed Parameters from Data and HiddenVector
as follows:
HiddenVector · Data
qA = !
HiddenVector · 1
!
( 1 HiddenVector) · Data
qB = ! ! .
( 1 HiddenVector) · 1
To make soft assignments, note that the assignment of outcomes to the two coins
can be represented by the binary responsibility matrix below. An occurrence of 1 in
the i-th position of the first row means that we conclude that coin A generated the i-th
sequence of flips, and an occurrence of 1 in the second row means that we conclude that
coin B generated the i-th sequence of flips:
0 1 1 0 1
HiddenMatrix
1 0 0 1 0
Thus, the first row of HiddenMatrix, denoted HiddenMatrix A , is just HiddenVector,

!
and the second row of HiddenMatrix, denoted HiddenMatrixB , is just 1 HiddenVector.
We can therefore rewrite the previous formulas for q A and q B in terms of HiddenMatrix:
HiddenMatrix A · Data
qA = !
HiddenMatrix A · 1
HiddenMatrixB · Data
qB = !
HiddenMatrixB · 1
99
CHAPTER 8
For the responsibility matrix in Figure 8.20, we can now recompute Parameters as
follows:
0.97 · 0.4 + 0.12 · 0.9 + 0.29 · 0.8 + 0.99 · 0.3 + 0.55 · 0.7 1.41
qA = = ⇡ 0.483
0.97 + 0.12 + 0.29 + 0.99 + 0.55 2.92
0.03 · 0.4 + 0.88 · 0.9 + 0.71 · 0.8 + 0.01 · 0.3 + 0.45 · 0.7 1.69
qB = = ⇡ 0.813
0.03 + 0.88 + 0.71 + 0.01 + 0.45 2.08
STOP and Think: Notice that the soft parameter choices q A = 0.483 and q B =
0.813 are a little closer to each other than the hard parameter choices q A = 0.467
and q B = 0.85. Why do you think that this is the case?
In general, the transition
(Data, HiddenMatrix, ?) ! Parameters
is called the M-step.
The expectation maximization algorithm
The expectation maximization algorithm starts with a random choice of Parameters.

It then alternates between the E-step, in which we compute a responsibility matrix
HiddenMatrix for Data given Parameters:
(Data, ?, Parameters) ! HiddenMatrix
and the M-step, in which we re-estimate Parameters using HiddenMatrix:
(Data, HiddenMatrix, ?) ! Parameters
EXERCISE BREAK: Carry out a few more steps of the expectation maximization
algorithm for the data in Figure 8.20. When should we stop the algorithm?
Soft k-Means Clustering
Applying expectation maximization to clustering
We are now ready to use the expectation maximization algorithm to modify the Lloyd al-
gorithm into a soft k-means clustering algorithm. This algorithm starts from randomly
chosen centers and iterates the following two steps:
100
• Centers to Soft Clusters (E-step): After centers have been selected, assign each
data point a “responsibility” for each cluster, where higher responsibilities corre-
spond to stronger cluster membership.
• Soft Clusters to Centers (M-step): After data points have been assigned to soft
clusters, compute new centers.
Centers to soft clusters
We begin with the “Centers to Soft Clusters” step. We have already used the term
“center of gravity” when computing centers; if we think about the centers as stars
and the data points as planets, then the closer a point is to a center, the stronger that
center’s “pull” should be on the point. Given k centers Centers = ( x1 , . . . xk ) and n
points Data = (Data1 , . . . , Datan ), we therefore need to construct a k ⇥ n responsibility
matrix HiddenMatrix for which HiddenMatrixi,j is the pull of center i on data point j. This
pull can be computed according to the Newtonian inverse-square law of gravitation,
1/d(Data j , xi )2
HiddenMatrixi,j = .
Âall centers xi 1/d(Data j , xi )2
Unfortunately for Newton fans, the following partition function from statistical
physics often works better in practice:
b·d(Data j , xi )
e
HiddenMatrixi,j = b·d(Data j , xi )
Âall centers xi e
In this formula, e is the base of the natural logarithm (e ⇡ 2.718), and b is a parameter
reflecting the amount of flexibility in our soft assignment and called — appropriately
enough — the stiffness parameter. Figure 8.21 illustrates different approaches for
computing HiddenMatrix when Data represents points in one-dimensional space.
STOP and Think: How does the assignment of the points in Figure 8.21 to soft
clusters change as b ! • or as b ! •? What about the case that b = 0?
EXERCISE BREAK: Compute HiddenMatrix using the Newtonian inverse-square

law for the three centers and eight data points shown in Figure 8.10 (left).
101
CHAPTER 8
-3 -2 -1 0 1 2 3
0.992 0.988 0.500 0.012 0.008

0.008 0.012 0.500 0.988 0.992
0.924 0.881 0.500 0.119 0.076

0.076 0.119 0.500 0.881 0.924
0.993 0.982 0.500 0.118 0.007

0.007 0.018 0.500 0.982 0.993
FIGURE 8.21 (Top) Five one-dimensional points Data = (-3, -2, 0, +2, +3) with two
centers (shown in blue and red) Centers = {-2.5, +2.5}. (Bottom) Three versions of
HiddenMatrix constructed for Data and Centers, using the Newtonian inverse-square
law (first matrix) and the partition function with stiffness b= 0.5 (second matrix), and
b= 1 (third matrix).
Soft clusters to centers
When we implemented the M-step for coin flipping, we obtained the following formulas
for q A and q B :
HiddenMatrix A · Data
qA = !
HiddenMatrix A · 1
HiddenMatrixB · Data
qB = !
HiddenMatrixB · 1
In soft k-means clustering, if we let HiddenMatrixi denote the i-th row of HiddenMatrix,
then we can update center xi using an analogue of the above formulas. Specifically, we
will define the j-th coordinate of center xi , denoted xi,j , as
HiddenMatrixi · Data j
xi,j = !
HiddenMatrixi · 1
Here, Data j is the n-dimensional vector holding the j-th coordinates of the n points in
Data. The updated center xi is called a weighted center of gravity of the points Data.
102
Computing weighted centers of gravity for the HiddenMatrix at the bottom of Fig-
ure 8.21 produces the following updated centers:
0.993 · ( 3) + 0.982 · ( 2) + 0.500 · (0) + 0.018 · (2) + 0.007 · (3)

x1 = = 1.955
0.993 + 0.982 + 0.500 + 0.018 + 0.007
0.007 · ( 3) + 0.018 · ( 2) + 0.500 · (0) + 0.982 · (2) + 0.993 · (3)
x2 = = 1.955
0.007 + 0.018 + 0.500 + 0.982 + 0.993
You are now ready to implement the expectation maximization algorithm for soft k- 8D
means clustering.
STOP and Think: Does the soft k-means clustering algorithm terminate? If not,
how would you modify it to ensure that it does not run forever?
EXERCISE BREAK: Recompute centers using the data points from the exercise
on page 100 along with the responsibility matrix obtained as the result of this
algorithm.
EXERCISE BREAK: Apply soft k-means clustering to the abridged yeast diauxic
shift gene expression data, and compare the results with those of the Lloyd
algorithm.
Hierarchical Clustering
Introduction to distance-based clustering
In Chapter 7, we discussed two approaches to evolutionary tree reconstruction with

different strengths and weaknesses: distance-based algorithms (including the neighbor-
joining algorithm), and alignment-based algorithms (including the algorithm for the
Small Parsimony Problem). Similarly, biologists do not always analyze the n ⇥ m gene
expression matrix directly. Instead, they sometimes first transform this matrix into
an n ⇥ n distance matrix D, where Di,j indicates the distance between the expression
vectors for genes i and j (Figure 8.22 (top right)). In this section, we will see how to
use a distance matrix to partition genes into clusters (see DETOUR: Transforming an PAGE 114
Expression Matrix into a Distance/Similarity Matrix for more details).
In previous sections, we assumed that we were working with a fixed number of
clusters k. But in practice, clusters often have subclusters, which have subsubclusters,
103
CHAPTER 8
1 hr 2 hr 3 hr g1 g2 g3 g4 g5 g6 g7 g8 g9 g10
g1 10.0 8.0 10.0 g1 0.0 8.1 9.2 7.7 9.3 2.3 5.1 10.2 6.1 7.0
g2 10.0 0.0 9.0 g2 8.1 0.0 12.0 0.9 12.0 9.5 10.1 12.8 2.0 1.0
g3 4.0 8.5 3.0 g3 9.2 12.0 0.0 11.2 0.7 11.1 8.1 1.1 10.5 11.5
g4 9.5 0.5 8.5 g4 7.7 0.9 11.2 0.0 11.2 9.2 9.5 12.0 1.6 1.1
g5 4.5 8.5 2.5 g5 9.3 12.0 0.7 11.2 0.0 11.2 8.5 1.0 10.6 11.6
g6 10.5 9.0 12.0 g6 2.3 9.5 11.1 9.2 11.2 0.0 5.6 12.1 7.7 8.5
g7 5.0 8.5 11.0 g7 5.1 10.1 8.1 9.5 8.5 5.6 0.0 9.1 8.3 9.3
g8 3.7 8.7 2.0 g8 10.2 12.8 1.1 12.0 1.0 12.1 9.1 0.0 11.4 12.4
g9 9.7 2.0 9.0 g9 6.1 2.0 10.5 1.6 10.6 7.7 8.3 11.4 0.0 1.1
g10 10.2 1.0 9.2 g10 7.0 1.0 11.5 1.1 11.6 8.5 9.3 12.4 1.1 0.0
g7
g6
g2 g10 g9 g1
g4
g3
g5 g8
g3 g5 g8 g7 g1 g6 g10 g2 g4 g9
FIGURE 8.22 A toy gene expression matrix of ten genes measured at three time points
(top left), the distance matrix based on Euclidean distance (top right), the gene expression
vectors as points in three-dimensional space (bottom left), and the tree produced from
the distance matrix by the hierarchical clustering algorithm (bottom right). Leaves
correspond to genes; internal nodes correspond to clusters of genes.
and so on. To capture this cluster stratification, the hierarchical clustering algorithm
uses an n ⇥ n distance matrix D to organize n data points into a tree (Figure 8.22 (bottom
right)). As shown in Figure 8.23, a horizontal line crossing the tree in i places divides
the n genes into i clusters.
EXERCISE BREAK: Figure 8.23 illustrates two ways of clustering the data from
Figure 8.22 using a tree. Find the remaining eight ways of clustering these data
using the same tree.
104
g7
g6
g2 g10 g9 g1
g4
g3
g5 g8
g3 g5 g8 g7 g1 g6 g10 g2 g4 g9
g7
g6
g2 g10 g9 g1
g4
g3
g5 g8
g3 g5 g8 g7 g1 g6 g10 g2 g4 g9
FIGURE 8.23 A tree with n leaves imposes n different ways of partitioning the data into
clusters. (Top) The horizontal line through the tree (left) crosses the data in four places
and partitions the data into four clusters (right). (Bottom) The same tree with a different
horizontal line (left) partitions the data into six clusters (right).
105
CHAPTER 8
Inferring clusters from a tree
H IERARCHICAL C LUSTERING, whose pseudocode is shown below, progressively gen-

erates n different partitions of the underlying data into clusters, all represented by
a tree in which each node is labeled by a cluster of genes. The first partition has n
single-element clusters represented by the leaves of the tree, with each element forming
its own cluster. The second partition merges the two “closest” clusters into a single
cluster consisting of two elements. In general, the i-th partition merges the two closest
clusters from the (i - 1)-th partition and has n i + 1 clusters. We hope this algorithm
looks familiar — it is UPGMA (from Chapter 7) in disguise.
HIERARCHICALCLUSTERING(D, n)
Clusters n single-element clusters labeled 1, . . . , n
construct a graph T with n isolated nodes labeled by single elements 1, . . . , n
while there is more than one cluster
find the two closest clusters Ci and Cj (break ties arbitrarily)
merge Ci and Cj into a new cluster Cnew with |Ci | + |Cj | elements
add a new node labeled by cluster Cnew to T
8E connect node Cnew to Ci and Cj by directed edges
remove the rows and columns of D corresponding to Ci and Cj
remove Ci and Cj from Clusters
add a row/column to D for Cnew by computing D(Cnew , C) for each C in Clusters
add Cnew to Clusters
root the node in T corresponding to the remaining cluster
return T
Note that we have not yet defined how H IERARCHICAL C LUSTERING computes the
distance D (Cnew , C ) between a newly formed cluster Cnew and each old cluster C. In
practice, clustering algorithms vary in how they compute these distances, with results
that can vary greatly. One commonly used approach (Figure 8.24) defines the distance
between clusters C1 and C2 as the smallest distance between any pair of elements from
these clusters,
Dmin (C1 , C2 ) = min Di,j .

all points i in cluster C1 , all points j in cluster C2
The distance function that we encountered with UPGMA uses the average distance
between elements in two clusters,
106
Âall points i in cluster C1 Âall points j in cluster C2 Di,j

Davg (C1 , C2 ) = .
|C1 | · |C2 |
g1 g2 g3 g4 g5 g6 g7 g8 g9 g10
g1 0.0 8.1 9.2 7.7 9.3 2.3 5.1 10.2 6.1 7.0
g2 8.1 0.0 12.0 0.9 12.0 9.5 10.1 12.8 2.0 1.0
g3 9.2 12.0 0.0 11.2 0.7 11.1 8.1 1.1 10.5 11.5 {g3, g5}
g4 7.7 0.9 11.2 0.0 11.2 9.2 9.5 12.0 1.6 1.1
g5 9.3 12.0 0.7 11.2 0.0 11.2 8.5 1.0 10.6 11.6
g6 2.3 9.5 11.1 9.2 11.2 0.0 5.6 12.1 7.7 8.5
g7 5.1 10.1 8.1 9.5 8.5 5.6 0.0 9.1 8.3 9.3 g3 g5 g8 g7 g1 g6 g10 g2 g4 g9
g8 10.2 12.8 1.1 12.0 1.0 12.1 9.1 0.0 11.4 12.4
g9 6.1 2.0 10.5 1.6 10.6 7.7 8.3 11.4 0.0 1.1
g10 7.0 1.0 11.5 1.1 11.6 8.5 9.3 12.4 1.1 0.0
g1 g2 g3 , g5 g4 g6 g7 g8 g9 g10
g1 0.0 8.1 9.2 7.7 2.3 5.1 10.2 6.1 7.0
g2 8.1 0.0 12.0 0.9 9.5 10.1 12.8 2.0 1.0 {g2, g4}
g3 , g5 9.2 12.0 0.0 11.2 11.1 8.1 1.0 10.5 11.5
g4 7.7 0.9 11.2 0.0 9.2 9.5 12.0 1.6 1.1 {g3, g5}
g6 2.3 9.5 11.1 9.2 0.0 5.6 12.1 7.7 8.5
g7 5.1 10.1 8.1 9.5 5.6 0.0 9.1 8.3 9.3
g8 10.2 12.8 1.0 12.0 12.1 9.1 0.0 11.4 12.4 g3 g5 g8 g7 g1 g6 g10 g2 g4 g9
g9 6.1 2.0 10.5 1.6 7.7 8.3 11.4 0.0 1.1
g10 7.0 1.0 11.5 1.1 8.5 9.3 12.4 1.1 0.0
g1 g2 , g4 g3 , g5 g6 g7 g8 g9 g10
g1 0.0 7.7 9.2 2.3 5.1 10.2 6.1 7.0 {g3, g5, g8}
g2 , g4 7.7 0.0 11.2 9.2 9.5 12.0 1.6 1.0
g3 , g5 9.2 11.2 0.0 11.1 8.1 1.0 10.5 11.5 {g2, g4}
g6 2.3 9.2 11.1 0.0 5.6 12.1 7.7 8.5
g7 5.1 9.5 8.1 5.6 0.0 9.1 8.3 9.3
g8 10.2 12.0 1.0 12.1 9.1 0.0 11.4 12.4
g9 6.1 1.6 10.5 7.7 8.3 11.4 0.0 1.1
g10 7.0 1.0 11.5 8.5 9.3 12.4 1.1 0.0 g3 g5 g8 g7 g1 g6 g10 g2 g4 g9
FIGURE 8.24 HIERARCHICALCLUSTERING in action. (Top left) The distance matrix from
Figure 8.22 (top left), with its minimum element shown in red, corresponding to
genes g3 and g5 . (Top right) Merging the single-element clusters containing g3 and g5 .
(Middle left) The updated distance matrix after computing Dmin for the new cluster
with respect to each other (single-element) cluster, with its minimum element shown in
red. (Middle right) Merging the two clusters corresponding to the minimum element.
(Bottom) Updating the distance matrix (left) and merging two additional clusters (right).
Subsequent steps will reconstruct the tree from Figure 8.22.
EXERCISE BREAK: Apply H IERARCHICAL C LUSTERING to the distance matrix

in Figure 8.22 using Davg instead of Dmin .
107
CHAPTER 8
EXERCISE BREAK: Apply H IERARCHICAL C LUSTERING (with Davg ) to the

abridged 230-gene yeast dataset, and partition this dataset into six clusters. Do
you expect these clusters to be roughly the same as the clusters shown in Fig-
ure 8.14? If not, should we be concerned?
Analyzing the diauxic shift with hierarchical clustering
Figure 8.25 visualizes expression vectors for each of the six clusters obtained after
applying H IERARCHICAL C LUSTERING (using Davg ) to the yeast dataset.

4 4 4
4
4
4
2 2 2
2
2
2
0 0 0
0
0
-2 -2 -2
2
2
-4 -4 -4
4

4 4 4
4
2 2 2
2
0 0 0
0
0
-2 -2 -2
2
2
-4 -4 -4
4
4
FIGURE 8.25 Applying HIERARCHICALCLUSTERING to the yeast dataset results in six

clusters with 22, 20, 87, 2, 95, and 4 genes, respectively.
STOP and Think: H IERARCHICAL C LUSTERING and the Lloyd algorithm (Fig-
ure 8.12) have produced different clusters. Should we be concerned?
EXERCISE BREAK: Implement H IERARCHICAL C LUSTERING (with Dmin

rather than Davg ) and apply it to partition the abridged yeast gene expression
dataset into six clusters. How does the result differ from Figure 8.25?
108
Biologists are not discouraged by the fact that different clustering approaches may
produce different clusters because applying these clustering algorithms is often just the
first step on the road to discovery (see DETOUR: Clustering and Corrupted Cliques PAGE 115
for yet another clustering approach). For this reason, gene expression studies are
typically followed by experimental work to confirm that derived clusters make sense
biologically. Once clusters have been generated, further research often focuses on
specific genes within these clusters.
For example, each of the clusters in Figure 8.25 can be further analyzed to reveal
sub-clusters of genes with even more pronounced expression profiles than the genes in
the entire cluster. In particular, cluster 1 contains seven genes exhibiting rather slight
changes during the first six checkpoints and a surge in gene expression at the final
checkpoint. Biologists discovered that six of these seven genes have the carbon source
response element (CSRE) regulatory motif with consensus sequence CATTCATCCG in
their upstream regions. Further analysis of the entire yeast genome revealed that only
four other yeast genes have this motif in their upstream region, suggesting that it was a
good idea to group these six genes into a sub-cluster within cluster 1.
The more important issue, however, is to understand why these six genes are related.
Yeast prefers using glucose as an energy source compared to other compounds like
ethanol, and so in the presence of glucose, the transcription of genes responsible for
metabolizing these less tasty compounds is repressed. Researchers have thus concluded
that the CSRE motif somehow helps yeast sense the presence of glucose and activates
the six genes in question when the organism runs out of glucose, thus serving as an
important component of the diauxic shift.
Finally, if you believe that we have exhausted every possible avenue of clustering,
take another look at Figure 8.16 (top). Although H IERARCHICAL C LUSTERING with
Dmin is able to find the clusters on the left and middle, none of the clustering algorithms
we have encountered can produce the clusters on the right. Clustering seems like a
straightforward problem in part because the human eye is so adept at grouping points
into shapes. After all, computer vision researchers are still trying to teach computers to
mimic our visual experience of the world, which is the outcome of millions of years of
evolution.
Epilogue: Clustering Tumor Samples
As we mentioned earlier, gene expression analysis has a wide variety of applications,

including cancer studies. In 1999, Uri Alon analyzed gene expression data for 2,000
109
CHAPTER 8
genes from 40 colon tumor tissues and compared them with data from colon tissues
belonging to 20 healthy individuals. We can represent his data as a 2,000 ⇥ 60 gene
expression matrix, where the first 40 columns describe tumor samples and the last 20
columns describe normal samples.
Now, suppose you performed a gene expression experiment with a colon sample
from a new patient, corresponding to a 61st column in an augmented gene expression
matrix. Your goal is to predict whether this patient has a colon tumor. Since the partition
of tissues into two clusters (tumor vs. healthy) is known in advance, it may seem that
classifying the sample from a new patient is easy. Indeed, since each patient corresponds
to a point in 2,000-dimensional space, we can compute the center of gravity of these
points for the tumor sample and for the healthy sample. Afterwards, we can simply
check which of the two centers of gravity is closer to the new tissue.
Alternatively, we could perform a blind analysis, pretending that we do not already
know the classification of samples into cancerous vs. healthy, and analyze the resulting
2,000 ⇥ 61 expression matrix to divide the 61 samples into two clusters. If we obtain a
cluster consisting predominantly of cancer tissues, this cluster may help us diagnose
colon cancer.
CHALLENGE PROBLEM: These approaches may seem straightforward, but

it is unlikely that either of them will reliably diagnose the new patient. Why do
you think this is the case? Given Alon’s 2,000 ⇥ 60 gene expression matrix and
gene data from a new patient, derive a superior approach to evaluate whether
this patient is likely to have a colon tumor.
110
Detours
Whole genome duplication or a series of duplications?
WGDs are quickly followed by massive gene loss and rearrangements, making it
difficult to reconstruct the pre-duplicated genome. Indeed, as we mentioned in the
main text, only 13% of genes have duplicates in modern-day S. cerevisiae. How, then,
can we argue that S. cerevisiae has indeed undergone a WGD instead of a sequence of
smaller duplications?
In 2004, Manolis Kellis analyzed K. waltii, a related yeast species. By aligning synteny
blocks from K. waltii and S. cerevisiae, he discovered that nearly every synteny block of
K. waltii aligns to two regions of S. cerevisiae. Because very few genes in the duplicated
S. cerevisiae blocks occurred in both blocks (Figure 8.26), Kellis argued that there was
indeed a WGD during yeast evolution.
S. cerevisiae copy 1
K. waltii
S. cerevisiae copy 2
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
FIGURE 8.26 Two synteny blocks from S. cerevisiae (whose respective genes are shown
in blue and green) aligned against a synteny block of K. waltii (whose genes are shown
in purple). Although only three of the sixteen genes in this synteny block have two
copies in S. cerevisiae, every gene in the K. waltii synteny block has a copy in at least
one of the two synteny blocks in S. cerevisiae. In fact, most synteny blocks in K. waltii
display this phenomenon, thus suggesting that there was indeed a WGD in S. cerevisiae.
STOP and Think: Although Manolis Kellis argued in 2004 that Figure 8.26
provides evidence for a WGD, three years later, Gustavo Caetano-Anollés raised
doubts about Kellis’s conclusion. Can you devise an alternative explanation for
Figure 8.26 and propose a different evolutionary scenario that does not require a
WGD?
Measuring gene expression
In the main text, we mentioned that we have encountered three technologies that
could be used to measure gene expression. First, in a mass spectrometry experiment
(Chapter 4), biologists generate a set of spectra and match them against a proteome.
111
CHAPTER 8
The number of spectra matching peptides from a given protein offers a proxy for the
expression level of this protein. To estimate the protein expression from this proxy,
bioinformaticians must account for the varying length of proteins, poor fragmentation
of some peptides (leading to difficulties in their identification), and other practical
concerns.
Second, in an RNA sequencing experiment, we generate reads from a transcrip-
tome, or all RNA transcripts present in a cell. By estimating the quantity of each
protein-coding RNA transcript in a sample, we obtain a proxy for the expression of
the resulting protein. A number of processes affect protein production in the cell in
addition to transcription, such as translation, post-translational modifications, and
protein degradation. These additional factors can muddle the correlation between the
quantity of a transcript and the expression of its corresponding protein.
Third, we can use DNA arrays (Chapter 2) carrying probes (k-mers) aimed at each
gene in a species of interest. Each probe is characterized by an intensity, which offers a
proxy for the number of transcripts of a given gene present in a sample. A deficiency
of DNA arrays is that they only target the identification of known transcripts and
often fail to evaluate unknown transcripts. For example, many cancers are caused
by rare mutations and would go undetected when using DNA arrays. As a result,
RNA sequencing is often more attractive in cancer studies, and it is now the dominant
technology for analyzing gene expression.
Microarrays
The microarrays that DeRisi used to study the diauxic shift were manufactured as fol-
lows. After capturing many RNA transcripts expressed in yeast cells, DeRisi converted
each RNA transcript to complementary DNA (cDNA) using an enzyme called reverse
transcriptase, and spotted these cDNAs on a glass slide. He then hybridized the cDNA
against fluorescently labeled RNA from a sample of interest in order to measure the
expression levels of various yeast genes.
The amount of cDNA printed on each spot of the microarray can vary greatly,
a complication that DeRisi needed to address in order to ensure that fluorescence
intensities could be compared across spots and across arrays. He therefore hybridized
two samples corresponding to two different timestamps to each array (Figure 8.27). He
then labeled the samples with different colors of fluorescent dyes so that the samples
could be distinguished by image-processing software.
The expression values obtained from a microarray are represented as the ratio of
fluorescent intensities of the two samples. Thus, expression is measured as relative
112
changes in the expression of individual genes between samples and time points. For
example, if a gene’s expression value is 2, then this gene’s expression is twice as large
in the first sample; if the expression value is 1/3, then the expression is three times as
large in the second sample. Following DeRisi, researchers commonly take the logarithm
of these expression ratios to generate expression matrices.
Sample 1 Sample 2
RNA Isolation
mRNA mRNA
Reverse
Transcriptase
Labeling
cDNA cDNA
“Red Fluorescent” Targets “Green Fluorescent” Targets
Combine Targets
Hybridize to
microarray
FIGURE 8.27 A microarray against which two samples of fluorescently labeled RNA
(red and green) has been hybridized.
Proof of the Center of Gravity Theorem
Note that when k = 1, the k-Means Clustering Problem is equivalent to finding a single
center point x that minimizes the sum of squared distances from x to points in Data.
Our goal is to show that the center of gravity of a set of points Data is the unique
point that minimizes D ISTORTION (Data, x ) over all possible centers x. Since the squared
Euclidean distance between DataPoint = (DataPoint1 , . . . , DataPointm ) and center x =
113
CHAPTER 8
( x1 , . . . , xm ) is equal to Â1 jm (DataPoint j x j )2 , we have that
1
D ISTORTION (Data, x ) = Â d(DataPoint, x )2
n all points DataPoint in Data
m
1
= Â Â (DataPointj x j )2
n all points DataPoint in Data j =1
m
1
= Â Â (DataPoint j x j )2 .
n j=1 all points DataPoint in Data
The last line of this formula implies that we can independently minimize D ISTORTION (Data, x )
in each of m dimensions by minimizing each of the m expressions
Â (DataPoint j x j )2 .
all points DataPoint in Data
Each of these expressions is a concave-up quadratic function of a single variable x j .

Thus, we can find the minimum of this function by finding where its derivative is equal
to zero:
Â 2 · (DataPoint j xj ) = 0 .
all points DataPoint in Data
The only solution of this equation is given by
1
xj = Â
n all points DataPoint
DataPoint j ,
in Data
implying that the center’s j-th coordinate is the mean value of the j-th coordinates of
the data points. In other words, the unique solution of the k-Means Clustering Problem
for k = 1 is simply the center of gravity of all data points.
Transforming an expression matrix into a distance/similarity matrix
There are many ways to quantify the similarity between expression vectors x =
( x1 , . . . , xm ) and y = (y, . . . , ym ). One possibility is the dot product, Âi=1,m xi · yi .
Another is the Pearson correlation coefficient P EARSON C ORRELATION ( x, y), where
Âim=1 ( xi µ( x )) · (yi µ(y))

P EARSON C ORRELATION ( x, y) = q .
Âim=1 ( xi µ( x ))2 · Âim=1 ( yi µ(y))2
114
In the above formula, µ( x ) denotes the means of all coordinates of vector x.
STOP and Think: Given a vector x, which vectors y maximize and minimize
P EARSON C ORRELATION ( x, y)?
The Pearson correlation coefficient varies between 1 and 1, where 1 indicates total
negative correlation, 0 indicates no correlation, and 1 indicates total positive correla-
tion. Based on the Pearson correlation coefficient, we can define the Pearson distance
between vectors x and y as
P EARSON D ISTANCE ( x, y) = 1 P EARSON C ORRELATION ( x, y) .
EXERCISE BREAK: Compute the Pearson correlation coefficient for the follow-
ing pairs of vectors:
1. (cos a, sin a) and (sin a, cos a) for an arbitrary value of a;

p p
2. ( 0.75, 0.5) and ( 0.75, 0.5).
Clustering and corrupted cliques
In expression analysis studies, a similarity matrix R is often transformed into a simi-

larity graph G ( R, q ). The nodes of this graph represent genes, and an edge connects
genes i and j if and only if the similarity between them (Ri,j ) exceeds a threshold value q.
STOP and Think: Consider a clustering of genes that satisfies the Good Clus-
tering Principle: the similarity between any two genes within the same cluster
exceeds q, and the similarity between any two genes in different clusters is less
than q. What does the similarity graph G ( R, q ) look like for these genes?
If clusters satisfy the Good Clustering Principle, then there should be some value of
q such that each connected component of G ( R, q ) is a clique, or a graph in which
every pair of nodes are connected by an edge (Figure 8.28). In general, a graph whose
connected components are all cliques is called a clique graph.
Errors in expression data and the absence of a universal threshold q often result in
corrupted similarity graphs whose connected components are not cliques (Figure 8.29).
Either genes from the same cluster may have a similarity value falling below q, thus
removing edges from a clique, or genes from different clusters may have a similarity
115
CHAPTER 8
value exceeding q, thus adding edges between different cliques. This observation leads
us to ask how to transform a corrupted similarity graph into a clique graph using the
smallest number of edge additions and deletions.
FIGURE 8.28 A clique graph consisting of three cliques.
g1 g10 g1 g10
g6 g9 g6 g9
g7 g2 g7 g2
g8 g4 g8 g4
g5 g3 g5 g3
FIGURE 8.29 (Left) One possible similarity graph for the genes from Figure 8.22. (Right)
The similarity graph can be transformed into a clique graph (right) by removing edges
(g1 , g10 ) and (g1 , g9 ).
Corrupted Cliques Problem:

Find the minimum number of edges that need to be added or deleted to transform a graph
into a clique graph.
Input: A graph.
Output: The minimum number of edge additions and deletions that trans-
form this graph into a clique graph.
116
The Corrupted Cliques Problem is difficult to solve exactly, so some heuristics have
been proposed. The Cluster Affinity Search Technique algorithm (CAST), described
below, performs remarkably well at clustering gene expression data.
Define the similarity between gene i and cluster C as the average similarity between
i and all genes in C:
Ri,j
Ri, C = Â |C |
all elements j in cluster C
Given a threshold q, a gene i is q-close to cluster C if Ri,C > q and q-distant from C
otherwise. A cluster is called consistent if all genes in C are q-close to C and all genes
not in C are q-distant from C. CAST uses the similarity graph and the threshold q to
iteratively find consistent clusters by starting with a single-element cluster C and then
adding the “closest” gene not in C and removing the “most distant” gene in C. After a
consistent cluster is found, all nodes in cluster C are removed from the similarity graph,
and CAST iterates over the resulting smaller graph.
CAST(R, j)
Graph G(R , j)
Clusters empty set
while Graph is nonempty
C a single-node cluster consisting of a node of maximal degree in Graph
while there exists a j-close gene i not in C or a j-distant gene i in C
find the nearest j-close gene i not in C and add it to C
find the farthest j-distant gene i in C and remove it from C
add C to the set Clusters
remove the nodes of C from Graph
return Clusters
EXERCISE BREAK: Implement CAST and use it to cluster the abridged gene
expression dataset.
117
CHAPTER 8
Bibliography Notes
The soft k-means algorithm for clustering, developed by Bezdek, 1981, is a variant of the
expectation maximization algorithm, which was first proposed by Ceppellini, Siniscalco,
and Smith, 1955 and rediscovered many times by various researchers. Do and Batzoglou,
2008 wrote an excellent primer on expectation maximization that inspired our discussion
of coin flipping. The Lloyd algorithm for k-means clustering was introduced by Lloyd,
1982. Arthur and Vassilvitskii, 2007 developed k-means++ initialization for k-means
clustering. The CAST algorithm was developed by Ben-Dor, Shamir, and Yakhini,
1999.
DeRisi, Iyer, and Brown, 1997 performed the first large-scale gene expression ex-
periment to analyze the diauxic shift (see Cristianini and Hahn, 2007 for an excellent
analysis of this experiment). Eisen et al., 1998 described the first applications of hier-
archical clustering to gene expression analyses. Alon et al., 1999 analyzed patterns of
gene expression in colon tumors.
Ohno, 1970 proposed the Whole Genome Duplication Model. Wolfe and Shields,
1997 provided the first convincing arguments in favor of a whole genome duplication
in yeast. Kellis, Birren, and Lander, 2004 provided further evidence for a whole genome
duplication by analyzing various yeast species. However, these arguments have not
convinced Martin et al., 2007, who published a rebuttal. Thomson et al., 2005 resurrected
the sequence of ancient alcohol dehydrogenases from yeast.
118
119
H O W D O W E L O C AT E D I S E A S E - C A U S I N G M U TAT I O N S ?
What Causes Ohdo Syndrome?
About 1% of babies are born with mental retardation, but this affliction remains poorly
understood because it can be caused by a variety of different genetic disorders. One of
these disorders is Ohdo syndrome, which causes an expressionless, “mask-like” face.
In 2011, biologists solved the genetic puzzle underlying Ohdo syndrome by discovering
a handful of mutations shared by multiple patients, which the researchers used to
identify a single protein-truncating mutation responsible for Ohdo syndrome.
The discovery of Ohdo syndrome’s root cause represents just one of many new
discoveries arising from the use of read mapping to study genetic disorders. In read
mapping, researchers compare sequenced DNA reads taken from an individual against
a reference human genome (see DETOUR: The Reference Human Genome) in order PAGE 170
to find which reads perfectly match the reference and which reads indicate mutations
of one nucleotide into another (single nucleotide polymorphisms, or SNPs). The ref-
erence genome is a gross simplification of species identity, since in addition to about
3 million SNPs (0.1% of the human genome), humans differ by genome rearrange-
ments, insertions, and deletions that can span thousands of nucleotides (see DETOUR: PAGE 170
Rearrangements, Insertions, and Deletions in Human Genomes). However, in this
chapter we will focus only on algorithms for finding SNPs.
But wait, you may say, why not use one of the algorithms that we already covered? After
all, we could always sequence the entire genome of an individual and then compare
it against the reference genome. However, sequencing methods are computationally
intensive and not perfect, as they often generate error-prone contigs. As a result, it
makes sense to map reads from an individual human to the reference human genome
to find out the differences.
To see why read mapping should be easier than genome assembly, let us return to
the analogy of a jigsaw puzzle, which is sold with a picture of the completed puzzle on
its box. This photo makes reconstructing the puzzle far easier; for a simple example, if
the completed puzzle shows a sun in a blue sky, then you can automatically move all of
the bright yellow pieces and all of the light blue pieces to the top of the puzzle.
Yet aside from genome sequencing, two other methods come to mind for read map-
ping. First, you could align each read to the reference genome (using a fitting alignment,
Chapter 5) to find the most similar region. Second, you could apply approximate pat-
tern matching algorithms to match each read one at a time against the reference genome.
STOP and Think: What computational challenges might arise from using these
methods to map millions of reads to a reference human genome?
121
CHAPTER 9
Both of these methods are guaranteed to solve the problem of mapping reads to a
reference genome, but their runtimes become a bottleneck when we scale to millions
of reads. Therefore, our goal in this chapter is to figure out how to use the reference
genome as a “photo on the box” shortcut to find SNPs.
Introduction to Multiple Pattern Matching
Recall from Chapter 3 that reads are typically a few hundred base pairs long. These
reads will form a collection of strings Patterns that we wish to match against a genome
Text. For each string in Patterns, we will first find all its exact matches as a substring
of Text (or conclude that it does not appear in Text). When hunting for the cause of a
genetic disorder, we can immediately eliminate from consideration areas of the reference
genome where exact matches occur. In the epilogue, we will generalize this problem to
find approximate matches, where single nucleotide substitutions in reads separate the
individual from the reference genome (or represent errors in reads).
Multiple Pattern Matching Problem:

Find all occurrences of a collection of patterns in a text.
Input: A string Text and a collection Patterns containing (shorter) strings.

Output: All starting positions in Text where a string from Patterns appears
as a substring.
A naive approach to the Multiple Pattern Matching Problem would attempt repeated ap-
plications of an algorithm for the (single) Pattern Matching Problem, which we encoun-
tered in Chapter 1. This algorithm, which we call B RUTE F ORCE PATTERN M ATCHING,
would slide each Pattern along Text, checking whether the substring starting at each posi-
tion of Text matches Pattern. Recall that the runtime of a naive algorithm for a single pat-
tern is O(|Text| · |Pattern|). Thus, the runtime of B RUTE F ORCE PATTERN M ATCHING
for the Multiple Pattern Matching Problem is O(|Text| · |Patterns|), where |Text| is the
length of Text and |Patterns| is the sum of the lengths of all strings in Patterns.
The problem with applying B RUTE F ORCE PATTERN M ATCHING to read mapping
is that |Text| and |Patterns| are both huge. In the case of the human genome (3 GB),
the total length of all reads may exceed 1 TB; as a result, any algorithm with runtime
O(|Text| · |Patterns|) will be too slow.
122
STOP and Think: The estimate O(|Text| · |Patterns|) presents the worst-case esti-
mate of the runtime for B RUTE F ORCE PATTERN M ATCHING. What is the average-
case estimate?
Herding Patterns into a Trie
Constructing a trie
The reason why the runtime of B RUTE F ORCE PATTERN M ATCHING is so high is that
each string in Patterns must traverse all of Text independently. If you think about Text
as a long road, then B RUTE F ORCE PATTERN M ATCHING is analogous to loading each
pattern into its own car when driving down Text, an inefficient strategy. Instead, our
goal is to herd the patterns onto a bus so that we only need to make one trip from the
beginning to the end of Text. In more formal terms, we would like to organize Patterns
into a data structure to prevent multiple passes down Text and to reduce the runtime.
To this end, we will consolidate Patterns into a directed acyclic graph called a trie
(pronounced “try”), which is written T RIE (Patterns) and has the following properties
(Figure 9.1).
• The trie has a single root node with indegree 0, denoted root.
• Each edge of T RIE (Patterns) is labeled with a letter of the alphabet.
• Edges leading out of a given node have distinct labels.
• Every string in Patterns is spelled out by concatenating the letters along some path
from the root downward.
• Every path from the root to a leaf, or node with outdegree 0, spells a string from
Patterns.
Trie Construction Problem:

Construct a trie from a collection of patterns.
Input: A collection of strings Patterns.

Output: T RIE (Patterns).
123
CHAPTER 9
root
a p
b n
n a a a
a d t n b n n
n e a d a
a n n a
s n a n
a a
FIGURE 9.1 The trie for the following collection of strings Patterns: "ananas", "and",
"antenna", "banana", "bandana", "nab", "nana", "pan".
The most obvious way to construct T RIE (Patterns) is by iteratively adding each
string from Patterns to the growing trie, as implemented by the following algorithm.
TRIECONSTRUCTION(Patterns)
Trie a graph consisting of a single node root
for each string Pattern in Patterns
currentNode root
for i 1 to |Pattern|
currentSymbol i-th symbol of Pattern
9A if there is an outgoing edge from currentNode with label currentSymbol
currentNode ending node of this edge
else
add a new node newNode to Trie
add a new edge from currentNode to newNode with label currentSymbol
currentNode newNode
return Trie
STOP and Think: How can we use the trie to solve the Multiple Pattern Matching
Problem?
124
Applying the trie to multiple pattern matching
Given a string Text and T RIE (Patterns), we can quickly check whether any string from
Patterns matches a prefix of Text. To do so, we start reading symbols from the begin-
ning of Text and see what string these symbols “spell” as we proceed along the path
downward from the root of the trie, as illustrated in Figure 9.2 (left). For each new
symbol in Text, if we encounter this symbol along an edge leading down from the
present node, then we continue along this edge; otherwise, we stop and conclude that
no string in Patterns matches a prefix of Text. If we make it all the way to a leaf, then
the pattern spelled out by this path matches a prefix of Text. This algorithm is called
P REFIX T RIE M ATCHING.
PREFIXTRIEMATCHING(Text, Trie)
symbol first letter of Text
v root of Trie
while forever
if v is a leaf in Trie
return the pattern spelled by the path from the root to v
else if there is an edge (v, w) in Trie labeled by symbol
symbol next letter of Text
v w
else
output “no matches found”
return
STOP and Think: For P REFIX T RIE M ATCHING to work, we have made a hidden
assumption that no string in Patterns is a prefix of another string in Patterns. How
can this algorithm be modified when Patterns is an arbitrary collection of strings?
Hint: consider adding "pantry" to the patterns in Figure 9.2 (left).
P REFIX T RIE M ATCHING finds whether any strings in Patterns match a prefix of Text.
To find whether any strings in Patterns match a substring of Text starting at position
k, we chop off the first k 1 symbols from Text and run P REFIX T RIE M ATCHING on
the shortened string. As a result, to solve the Multiple Pattern Matching Problem, we
simply iterate P REFIX T RIE M ATCHING |Text| times, chopping the first symbol off of
Text before each new iteration (Figure 9.2 (right)).
125
CHAPTER 9
root root
a p a p
b n b n
n a a a n a a a
a d t n b n n a d t n b n n
n e a d a n e a d a
a n n a a n n a
s n a n s n a n
a a a a
FIGURE 9.2 (Left) The pattern "pan" matches Text = "panamabananas" when starting
at the beginning of Text. (Right) No pattern match is found in TRIE (Patterns) for the
strings Patterns from Figure 9.1 when starting at the third symbol of "panamabananas".
TRIEMATCHING(Text, Trie)
while Text is nonempty
9B PREFIXTRIEMATCHING (Text, Trie)
remove first symbol from Text
We need |Patterns| steps to construct T RIE (Patterns), which contains at most |Patterns|
nodes. Each iteration of P REFIX T RIE M ATCHING takes at most |LongestPattern| steps,
where LongestPattern is the longest string in Patterns. T RIE M ATCHING makes |Text| total
calls to P REFIX T RIE M ATCHING, making the total number of steps equal to |Patterns| +
|Text| · |LongestPattern|. This runtime offers a huge speed-up compared to the |Text| ·
|Patterns| steps required by B RUTE F ORCE PATTERN M ATCHING. The Aho-Corasick
algorithm, developed in 1975, further reduces the number of steps required after con-
PAGE 170 structing the trie from O(|Text| · |LongestPattern|) steps to O(|Text|) steps (see DETOUR:
The Aho-Corasick Algorithm).
STOP and Think: Do you see any computational challenges with using
T RIE M ATCHING to solve the Multiple Pattern Matching Problem?
Although T RIE M ATCHING is fast, storing a trie consumes a lot of memory. Recall
that B RUTE F ORCE PATTERN M ATCHING works with a single read at a time, which
keeps the memory low because we only need to store the genome in memory. Yet
T RIE M ATCHING needs to store the entire trie in memory, which is proportional to
126
|Patterns|. Since a collection of reads for the human genome may consume upwards of
1 TB, the memory required to store the trie is prohibitive.
STOP and Think: How can we avoid multiple passes through the genome
without needing to consolidate all the reads into a huge data structure?
Preprocessing the Genome Instead
Introduction to suffix tries
Since storing T RIE (Patterns) requires so much memory, let’s process Text into a data
structure instead. Our goal is to compare each string in Patterns against Text without
needing to traverse Text from beginning to end. In more familiar terms, instead of
packing Patterns onto a bus and riding the long distance down Text, our new data
structure will be able to “teleport” each string in Patterns directly to its occurrences in
Text.
A suffix trie, denoted S UFFIX T RIE (Text), is the trie formed from all suffixes of Text
(Figure 9.3). From now on, we append the dollar-sign ("$") to Text in order to mark the
end of Text. We will also label each leaf of the resulting trie by the starting position of
the suffix whose path through the trie ends at this leaf (using 0-based indexing). This
way, when we arrive at a leaf, we will immediately know where this suffix came from
in Text.
STOP and Think: How can we use the suffix trie for pattern matching?
Using suffix tries for pattern matching
To match a single string Pattern to Text, note that if Pattern matches a substring of
Text starting at position i, then Pattern must also appear at the beginning of the suffix
of Text starting at position i. We can therefore determine whether Pattern occurs in
S UFFIX T RIE (Text) by starting at the root and spelling symbols of Pattern downward. If
we can find a path in the suffix trie spelling out Pattern, then we know that Pattern must
occur in Text (Figure 9.4). We can then iterate over all strings in Patterns.
STOP and Think: Figure 9.4 illustrates how to find the pattern "nanas" in
S UFFIX T RIE ("panamabananas$"), but it does not tell us where "nanas" occurs
in Text. How can we obtain this information?
127
CHAPTER 9
root
$ s
a m p
b n
13
b s a a a a $
m n
12
a a a $ n b m n s n
11
n b m n s a a a a $ a
10
a a a a $ n n b s m
9
n n b s a a a $ a
8
a a a $ s n n b
7
s n n $ a a a
6
$ a a s n n
5
s n $ a a
4
$ a s n
3
s $ a
2
$ s
1
$
0
FIGURE 9.3 SUFFIXTRIE ("panamabananas$"), with leaf labels (corresponding to start-

ing positions of suffixes) varying from 0 to 13.
To determine where Pattern appears in Text, assume first that Pattern matches Text at a
leaf of S UFFIX T RIE (Text). In this case, Pattern must appear in Text as a suffix, and we
can consult the label at that leaf to determine the starting position of the suffix. For
example, threading "nanas" into the suffix trie in Figure 9.4 shows that it matches the
suffix starting at position 8 of "panamabananas$".
If the path spelling out Pattern stops before a leaf at some node v of S UFFIX T RIE (Text),
then Pattern may occur more than once in Text. To locate these occurrences, follow all
paths from v down to the leaves of S UFFIX T RIE (Text), which will indicate all starting
128
root
$ s
a m p
b n
13
b s a a a a $
m n
12
a a a $ n b m n s n
11
10
a a a a $ n n b s m
9
n n b s a a a $ a
8
a a a $ s n n b
7
s n n $ a a a
6
$ a a s n n
5
s n $ a a
4
$ a s n
3
s $ a
2
$ s
1
$
0
FIGURE 9.4 Threading "antenna" through SUFFIXTRIE ("panamabananas$") fails to

find a match because no suffix of "panamabananas$" begins with "ant"; however,
threading "nana" through the suffix trie does find a match: "panamabananas$".
positions of Pattern in Text. For example, the pattern "ana" corresponds to a path in
S UFFIX T RIE ("panamabananas$") that can be extended to three different leaves with
labels 1, 7, and 9, corresponding to three occurrences of "ana": "panamabananas$",
"panamabananas$", and "panamabananas$".
STOP and Think: How much runtime and memory will it take to construct
S UFFIX T RIE (Text)?
129
CHAPTER 9
root
$ s
a m p
b n
13
b s a a a a $
m n
12
a a a $ n b m n s n
11
10
a a a a $ n n b s m
9
n n b s a a a $ a
8
a a a $ s n n b
7
s n n $ a a a
6
$ a a s n n
5
s n $ a a
4
$ a s n
3
s $ a
2
$ s
1
$
0
FIGURE 9.5 All paths starting with "ana" reveal the three occurrences of "ana" in
"panamabananas$" . Extending these paths to the leaves (shown in green) reveals that
the starting positions of these occurrences are 1, 7, and 9.
Recall that constructing T RIE (Patterns) required O(|Patterns|) runtime and memory.
Accordingly, the runtime and memory required to construct S UFFIX T RIE (Text) are both
equal to the combined length of all suffixes in Text. There are |Text| suffixes of Text,
ranging in length from 1 to |Text| and having total length |Text| · (|Text| + 1) 2 , which
is O |Text|2 . Thus, we need to reduce both the construction time and memory require-
ments of suffix tries to make them practical.
130
Suffix Trees
Let’s not give up hope on suffix tries. We can reduce the number of edges in the suffix
trie by combining the edges on any non-branching path into a single edge. We then label
this edge with the concatenation of symbols on the consolidated edges, as shown in
Figure 9.6. The resulting data structure is called a suffix tree, written S UFFIX T REE (Text).
To match a single Pattern to Text, we thread Pattern into S UFFIX T REE (Text) by the
same process used for a suffix trie. Similarly to the suffix trie, we can use the leaf labels
to find starting positions of successfully matched patterns.
EXERCISE BREAK: Prove that S UFFIX T REE (Text) has exactly |Text| + 1 leaves
and at most |Text| + 1 other nodes.
root
$
a m
13 a
b b n p s
a a a a $
n n n
a a a 12
b n s n n m
a m a $ a a a
n a s s n b
a b 11 $ $ a a
n a m s s n
a n 6 4 a $ $
n b a
s a n
n a a 8 10
$ m s s a
a n
a $ $ a s
s b $
5 $ n
a 7 9 a
n s 0
3 a $
n
a 2
s
$
1
FIGURE 9.6 SUFFIXTREE ("panamabananas$"), formed by compressing the edges on

non-branching paths in Figure 9.3.
Compared to a suffix trie, which may have a quadratic number of nodes in the length
of Text, the number of nodes in S UFFIX T REE (Text) is at most 2 · |Text|. Therefore, the
memory required for S UFFIX T REE (Text) is O(|Text|).
STOP and Think: Wait a second. Suffix trees seem like a cosmetic modification
of suffix tries. Since we still need to keep all concatenated edge labels in memory,
why should suffix trees be more memory-efficient than suffix tries?
131
CHAPTER 9
Suffix trees save memory because they do not need to store concatenated edge labels
from each non-branching path. For example, a suffix tree does not need ten bytes to
store the edge labeled "mabananas$" in Figure 9.6; instead, it suffices to store a pointer
to position 4 of "panamabananas$", as well as the length of "mabananas$". Furthermore,
suffix trees can be constructed in linear time, without having to first construct the suffix
trie! We will not ask you to implement this fast suffix tree construction algorithm
because it is quite complex.
Suffix Tree Construction Problem:

Construct the suffix tree of a string.
9C
Input: A string Text.
Output: S UFFIX T REE (Text).
CHARGING STATION (Constructing a Suffix Tree): A memory-inefficient

way of constructing the suffix tree is to first construct the suffix trie and then con-
PAGE
solidate each non-branching path into a single edge, storing its label in memory.
164 To implement a more memory-efficient solution, check out this Charging Station.
Although the suffix tree decreases memory requirements from O |Text|2 to O(|Text|),
on average it still requires about 20 times as much memory as Text. In the case of a
3 GB human genome, 60 GB of RAM is a huge improvement over the 1 TB that we
needed to work with T RIE (Patterns), but it still presents a memory challenge for most
machines. This reveals a dark secret of big-O notation, which is that it ignores constant
factors. For long strings such as the human genome, we will need to pay attention to
this constant factor, since the expression O(|Text|) applies to both an algorithm with
2 · |Text| memory and an algorithm with 1000 · |Text| memory.
Yet before seeing how we can further reduce the memory needed for multiple pat-
tern matching, we ask you to solve three problems for which suffix trees are useful.
Longest Repeat Problem:

Find the longest repeat in a string.
9D
Output: A longest substring of Text that appears in Text more than once.
132
Longest Shared Substring Problem:

Find the longest substring shared by two strings.
9E
Input: Strings Text1 and Text2 .
Output: The longest substring that occurs in both Text1 and Text2 .
Shortest Non-Shared Substring Problem:

Find the shortest substring of one string that does not appear in another string.
9F
Input: Strings Text1 and Text2 .
Output: The shortest substring of Text1 that does not appear in Text2 .
CHARGING STATION (Solving the Longest Shared Substring Problem):

One way of solving the Longest Shared Substring Problem is to construct two
suffix trees, one for Text1 and one for Text2 . Check out this Charging Station to PAGE
learn about a more elegant solution. 167
Suffix Arrays
Constructing a suffix array
In 1993, Udi Manber and Gene Myers introduced suffix arrays as a memory-efficient
alternative to suffix trees. To construct S UFFIX A RRAY (Text), we first sort all suffixes of
Text lexicographically, assuming that "$" comes first in the alphabet (Figure 9.7). The
suffix array is the list of starting positions of these sorted suffixes:
S UFFIX A RRAY ("panamabananas$") = [13, 5, 3, 1, 7, 9, 11, 6, 4, 2, 8, 10, 0, 12].
Suffix Array Construction Problem:

Construct the suffix array of a string.
9G
Output: S UFFIX A RRAY (Text).
133
CHAPTER 9
Sorted Suffixes Starting Positions

$ 13
abananas$ 5
amabananas$ 3
anamabananas$ 1
ananas$ 7
anas$ 9
as$ 11
bananas$ 6
mabananas$ 4
namabananas$ 2
nanas$ 8
nas$ 10
panamabananas$ 0
s$ 12
FIGURE 9.7 The sorted list of suffixes of Text = "panamabananas$", along with their
starting positions in Text; these starting positions form the suffix array of Text.
The Suffix Array Construction Problem can easily be solved after sorting all suffixes
of Text, but since even the fastest algorithms for sorting an array of n elements require
O(n · log n) comparisons, sorting all suffixes takes O(|Text| · log (|Text|)) comparisons.
However, there exists a faster algorithm that constructs suffix arrays in linear time and
requires only about a fifth as much memory as suffix trees, which knocks the 60 GB
memory requirement for the human genome down to 12 GB.
STOP and Think: Given a suffix tree, can you quickly transform it into a suffix
array? Given a suffix array, can you quickly transform it into a suffix tree?
As the preceding question suggests, suffix arrays and suffix trees are practically equiv-
alent; every algorithm using suffix trees can be translated into an algorithm using
PAGE 171 suffix arrays (see DETOUR: From Suffix Trees to Suffix Arrays), and vice-versa (see
DETOUR: From Suffix Arrays to Suffix Trees).
PAGE 173
Pattern matching with the suffix array
Once we have constructed the suffix array of a string Text, we can use it to quickly locate
every occurrence of a string Pattern in Text. First, recall that when pattern matching
with the suffix trie, we observed that all matches of Pattern in Text must occur at the
beginning of suffixes of Text. Second, note that after sorting the suffixes of Text, the
134
suffixes beginning with Pattern clump together. For example, in Figure 9.7, Pattern
= "ana" occurs at the beginning of the suffixes "anamabananas$", "ananas$", and
"anas$" of Text = "panamabananas$"; these suffixes occur in three consecutive rows
and correspond to the starting positions 1, 7, and 9 in Text.
The question is how to find these starting positions for an arbitrary string Pattern
without needing to store the sorted suffixes of Text. The following algorithm, called
PATTERN M ATCHING W ITH S UFFIX A RRAY, identifies the first and last index of the suf-
fix array corresponding to suffixes beginning with Pattern (these indices are denoted
first and last, respectively). PATTERN M ATCHING W ITH S UFFIX A RRAY offers a variation
of a general search technique called binary search that finds a data point in a sorted
collection of data by iteratively dividing the data in half and determining the half in
which the data point lies. See DETOUR: Binary Search for more details. PAGE 176
PATTERNMATCHINGWITHSUFFIXARRAY(Text, Pattern, SUFFIXARRAY)

minIndex 0
maxIndex |Text|
while minIndex < maxIndex
midIndex (minIndex + maxIndex)/2
if Pattern > suffix of Text starting at position SUFFIXARRAY (midIndex)
minIndex midIndex + 1
else
maxIndex midIndex
first minIndex
maxIndex |Text| 9H
while minIndex < maxIndex
if Pattern < suffix of Text starting at position SUFFIXARRAY (midIndex)
maxIndex midIndex
else
last maxIndex
if first > last
return "Pattern does not appear in Text"
else
return (first, last)
135
CHAPTER 9
The Burrows-Wheeler Transform
Genome compression
Suffix arrays have greatly reduced the memory required for efficient text searches, and
until the start of this century, they represented the state of the art in pattern matching.
Can we be so ambitious as to look for a data structure that would encode Text using
memory approximately equal to the length of Text while still enabling fast pattern
matching?
To answer this question, we will digress to consider the seemingly unrelated topic
of text compression. In one simple compression technique called run-length encod-
ing, we replace a run of k consecutive occurrences of symbol s with only two sym-
bols: k, followed by s. For example, run-length encoding would compress the string
TTTTTGGGAAAACCCCCCA into 5T3G4A6C1A.
Run-length encoding works well for strings having lots of long runs, but real
genomes do not have many runs. What they do have, as we saw in Chapter 3, are
repeats. It would therefore be nice if we could first manipulate the genome to convert
repeats into runs and then apply run-length encoding to the resulting string.
A naive way of creating runs in a string is to reorder the string’s symbols lexico-
graphically. For example, TACGTAACGATACGAT would become AAAAACCCGGGTTTT,
which we could then compress into 5A3C3G4T. This method would represent a 3 GB
human genome file using just four numbers.
STOP and Think: What is wrong with applying this compression method to
genomes?
Ordering a string’s symbols lexicographically is not suitable for compression because

many different strings will get compressed into the same string. For example, the
DNA strings GCATCATGCAT and ACTGACTACTG — as well as any string with the
same nucleotide counts — get reordered into AAACCCGGTTT. As a result, we cannot
decompress the compressed string, i.e., invert the compression operation to produce
the original string.
Constructing the Burrows-Wheeler transform
Let’s consider a different method of converting the repeats of a string into runs that was
proposed by Michael Burrows and David Wheeler in 1994. First, form all possible cyclic
rotations of Text; a cyclic rotation is defined by chopping off a suffix from the end of
136
Text and appending this suffix to the beginning of Text. Next — similarly to suffix arrays
— order all the cyclic rotations of Text lexicographically to form a |Text| ⇥ |Text| matrix of
symbols that we call the Burrows-Wheeler matrix and denote by M (Text) (Figure 9.8).
Cyclic Rotations M("panamabananas$")

panamabananas$ $ p a n a m a b a n a n a s
$panamabananas a b a n a n a s $ p a n a m
s$panamabanana a m a b a n a n a s $ p a n
as$panamabanan a n a m a b a n a n a s $ p
nas$panamabana a n a n a s $ p a n a m a b
anas$panamaban a n a s $ p a n a m a b a n
nanas$panamaba a s $ p a n a m a b a n a n
ananas$panamab b a n a n a s $ p a n a m a
bananas$panama m a b a n a n a s $ p a n a
abananas$panam n a m a b a n a n a s $ p a
mabananas$pana n a n a s $ p a n a m a b a
amabananas$pan n a s $ p a n a m a b a n a
namabananas$pa p a n a m a b a n a n a s $
anamabananas$p s $ p a n a m a b a n a n a
FIGURE 9.8 All cyclic rotations of "panamabananas$" (left) and the Burrows-
Wheeler matrix M("panamabananas$") of all lexicographically ordered cyclic rota-
tions (right). BWT("panamabananas$") is the last column of M("panamabananas$"):
"smnpbnnaaaaa$a".
Notice that the first column of M (Text) contains the symbols of Text ordered lexico-
graphically, which is just the naive rearrangement of Text that we already described.
In turn, the second column of M (Text) contains the second symbols of all cyclic rota-
tions of Text, and so it too represents a (different) rearrangement of symbols from Text.
The same reasoning applies to show that any column of M (Text) is some rearrange-
ment of the symbols of Text. We are interested in the last column of M (Text), called the
Burrows-Wheeler transform of Text, or BWT (Text), which is shown in red in Figure 9.8.
STOP and Think: We have seen that the first column of M (Text) cannot be
uniquely decompressed to yield Text. Do you think that some other column of
M (Text) can be inverted to yield Text?
137
CHAPTER 9
Burrows-Wheeler Transform Construction Problem:

Construct the Burrows-Wheeler transform of a string.
9I
Output: BWT (Text).
STOP and Think: Figure 9.8 suggests a simple algorithm for computing
BWT (Text) based on constructing M (Text). Can you construct BWT (Text) using
less memory given Text and S UFFIX A RRAY (Text)?
From repeats to runs
If we re-examine the Burrows-Wheeler transform in Figure 9.8, we immediately notice

that it has created the run "aaaaa" in BWT ("panamabananas$") = "smnpbnnaaaaa$a".
STOP and Think: Why do you think that the Burrows-Wheeler Transform pro-
duced this run?
Imagine that we take the Burrows-Wheeler transform of Watson and Crick’s 1953 paper
on the double helix structure of DNA. The word "and" is repeated often in English,
which means that when we form all possible cyclic rotations of the Watson & Crick
paper, we will witness a large number of rotations beginning with "and..." In turn,
we will observe many rotations that begin with "nd..." and end with "...a". When
all the cyclic rotations of Text are sorted lexicographically to form M (Text), all rows that
begin with "nd..." and end with "...a" will tend to clump together. As illustrated in
Figure 9.9, this clumping produces runs of "a" in the final column of M (Text), which
we know is BWT (Text).
The substring "ana" in "panamabananas$" plays the role of "and" in Watson and
Crick’s paper and explains three of the five occurrences of "a" in the repeat "aaaaa"
in BWT ("panamabananas$") = "smnpbnnaaaaa$a". When the Burrows-Wheeler trans-
form is applied to a genome, it converts the genome’s many repeats into runs. As we
already suggested, after applying the Burrows-Wheeler transform, we can apply an
additional compression method such as run-length encoding in order to further reduce
the memory.
138
nd Corey (1). They kindly made their manuscript availa ...... a

nd criticism, especially on interatomic distances. We ...... a
nd cytosine. The sequence of bases on a single chain d ...... a
nd experimentally (3,4) that the ratio of the amounts o ...... u
nd for this reason we shall not comment on it. We wish ...... a
nd guanine (purine) with cytosine (pyrimidine). In oth ...... a
nd ideas of Dr. M. H. F. Wilkins, Dr. R. E. Franklin ...... a
nd its water content is rather high. At lower water co ...... a
nd pyrimidine bases. The planes of the bases are perpe ...... a
nd stereochemical arguments. It has not escaped our no ...... a
nd that only specific pairs of bases can bond together ...... u
nd the atoms near it is close to Furberg’s ’standard co ...... a
nd the bases on the inside, linked together by hydrogen ...... a
nd the bases on the outside. In our opinion, this stru ...... a
nd the other a pyrimidine for bonding to occur. The hy ...... a
nd the phosphates on the outside. The configuration of ...... a
nd the ration of guanine to cytosine, are always very c ...... a
nd the same axis (see diagram). We have made the usual ...... u
nd their co-workers at King’s College, London. One of ...... a
FIGURE 9.9 A few consecutive rows selected from M(Text), where Text is Watson and
Crick’s 1953 paper on the double helix. Rows beginning with "nd..." often end with
"...a" because of the common occurrence of the word "and" in English, which causes
runs of "a" in BWT(Text).
EXERCISE BREAK: There is only one run of length at least 10 in the E. coli
genome. How many runs of length at least 10 do you find after applying the
Burrows-Wheeler transform to the E. coli genome?
Inverting the Burrows-Wheeler Transform
A first attempt at inverting the Burrows-Wheeler transform
Before we get ahead of ourselves, remember that compressing a genome does not count
for much if we cannot decompress it. In particular, if there exist a pair of genomes that
the Burrows-Wheeler transform compresses into the same string, then we will not be
able to decompress this string. But it turns out that the Burrows-Wheeler transform is
reversible!
STOP and Think: Can you find the (unique) string whose Burrows-
Wheeler transform is "enwvpeoseu$llt"? It could be "newtloveslupe$",
"elevenplustwo$", "unwellpesovet$", or something else entirely.
139
CHAPTER 9
Consider the toy example BWT (Text) = "ard$rcaaaabb". First, recall that the first
column of M (Text) is the lexicographic rearrangement of symbols in BWT (Text), i.e.,
"$aaaaabbcdrr". For convenience, we will use the terms FirstColumn and LastColumn
(i.e., BWT (Text)) when referring to the first and last columns of M (Text), respectively.
We know that the first row of M (Text) is the cyclic rotation of Text beginning with
"$", which occurs at the end of Text. Thus, if we determine the first row of M (Text),
then we can move the "$" to the end of this row and reproduce Text. But how do we
determine the remaining symbols in this first row, if all we know is FirstColumn and
LastColumn?
$ ? ? ? ? ? ? ? ? ? ? a
a ? ? ? ? ? ? ? ? ? ? r
a ? ? ? ? ? ? ? ? ? ? d
a ? ? ? ? ? ? ? ? ? ? $
a ? ? ? ? ? ? ? ? ? ? r
a ? ? ? ? ? ? ? ? ? ? c
b ? ? ? ? ? ? ? ? ? ? a
b ? ? ? ? ? ? ? ? ? ? a
c ? ? ? ? ? ? ? ? ? ? a
d ? ? ? ? ? ? ? ? ? ? a
r ? ? ? ? ? ? ? ? ? ? b
r ? ? ? ? ? ? ? ? ? ? b
STOP and Think: Using the first and last columns of the Burrows-Wheeler
matrix shown above, can you find the first symbol of Text?
Note that the first symbol in Text must follow "$" in any cyclic rotation of Text. Because
"$" occurs as the fourth symbol of LastColumn = "ard$rcaaaabb", we know that if
we walk one symbol to the right from the end of the fourth row of M (Text), then we
will “wrap around” and arrive at the fourth symbol of FirstColumn, which is "a" in
"$aaaaabbcdrr". Therefore, this "a" belongs in the first position of Text:
$ a ? ? ? ? ? ? ? ? ? a
a ? ? ? ? ? ? ? ? ? ? r
a ? ? ? ? ? ? ? ? ? ? d
a ? ? ? ? ? ? ? ? ? ? $
a ? ? ? ? ? ? ? ? ? ? r
a ? ? ? ? ? ? ? ? ? ? c
b ? ? ? ? ? ? ? ? ? ? a
b ? ? ? ? ? ? ? ? ? ? a
c ? ? ? ? ? ? ? ? ? ? a
d ? ? ? ? ? ? ? ? ? ? a
r ? ? ? ? ? ? ? ? ? ? b
r ? ? ? ? ? ? ? ? ? ? b
140
STOP and Think: Which symbol is hiding in the second position of Text?
Following the same logic of “wrapping around”, the next symbol of Text should be the
first symbol in a row of M (Text) that ends in "a". The only trouble is that five rows end
in "a", and we don’t know which of them is the correct one! If we guess that this "a" is
the seventh symbol of "ard$rcaaaabb", then we obtain "b" in the second position of
Text (Figure 9.10 (left)). On the other hand, if we guess that this "a" is the ninth symbol
of "ard$rcaaaabb", then we obtain "c" in the second position of Text (Figure 9.10 (mid-
dle)). Finally, if we guess that this "a" is the tenth symbol of "ard$rcaaaabb", then we
obtain "d" in the second position of Text (Figure 9.10 (right)).
$ a b ? ? ? ? ? ? ? ? a $ a c ? ? ? ? ? ? ? ? a $ a d ? ? ? ? ? ? ? ? a
a ? ? ? ? ? ? ? ? ? ? r a ? ? ? ? ? ? ? ? ? ? r a ? ? ? ? ? ? ? ? ? ? r
a ? ? ? ? ? ? ? ? ? ? d a ? ? ? ? ? ? ? ? ? ? d a ? ? ? ? ? ? ? ? ? ? d
a ? ? ? ? ? ? ? ? ? ? $ a ? ? ? ? ? ? ? ? ? ? $ a ? ? ? ? ? ? ? ? ? ? $
a ? ? ? ? ? ? ? ? ? ? r a ? ? ? ? ? ? ? ? ? ? r a ? ? ? ? ? ? ? ? ? ? r
a ? ? ? ? ? ? ? ? ? ? c a ? ? ? ? ? ? ? ? ? ? c a ? ? ? ? ? ? ? ? ? ? c
b ? ? ? ? ? ? ? ? ? ? a b ? ? ? ? ? ? ? ? ? ? a b ? ? ? ? ? ? ? ? ? ? a
b ? ? ? ? ? ? ? ? ? ? a b ? ? ? ? ? ? ? ? ? ? a b ? ? ? ? ? ? ? ? ? ? a
c ? ? ? ? ? ? ? ? ? ? a c ? ? ? ? ? ? ? ? ? ? a c ? ? ? ? ? ? ? ? ? ? a
d ? ? ? ? ? ? ? ? ? ? a d ? ? ? ? ? ? ? ? ? ? a d ? ? ? ? ? ? ? ? ? ? a
r ? ? ? ? ? ? ? ? ? ? b r ? ? ? ? ? ? ? ? ? ? b r ? ? ? ? ? ? ? ? ? ? b
r ? ? ? ? ? ? ? ? ? ? b r ? ? ? ? ? ? ? ? ? ? b r ? ? ? ? ? ? ? ? ? ? b
FIGURE 9.10 The three possibilities ("b", "c", or "d") for the third element of the first
row of M(Text) when BWT(Text) is "ard$rcaaaabb". One of these possibilities must
correspond to the second symbol of Text.
STOP and Think: How would you choose among "b", "c", and "d" for the
second symbol of Text?
The First-Last Property
To determine the remaining symbols of Text, we need to use a subtle property of M (Text)
that may seem completely unrelated to inverting the Burrows-Wheeler transform. Be-
low, we have indexed the occurrences of each symbol in FirstColumn with subscripts
according to their order of appearance in this column. When Text = "panamabananas$",
six instances of "a" appear in FirstColumn.
141
CHAPTER 9
$ p a n a m a b a n a n a s
a1 b a n a n a s $ p a n a m
a2 m a b a n a n a s $ p a n
a3 n a m a b a n a n a s $ p
a4 n a n a s $ p a n a m a b
a5 n a s $ p a n a m a b a n
a6 s $ p a n a m a b a n a n
b a n a n a s $ p a n a m a
m a b a n a n a s $ p a n a
n a m a b a n a n a s $ p a
n a n a s $ p a n a m a b a
n a s $ p a n a m a b a n a
p a n a m a b a n a n a s $
s $ p a n a m a b a n a n a
Consider "a1 " in FirstColumn, which occurs at the beginning of the cyclic rotation
"a1 bananas$panam". If we cyclically rotate this string, then we obtain "panama1 bananas$".
Thus, "a1 " in FirstColumn is actually the third occurrence of "a" in "panamabananas$".
We can now identify the positions of the other five instances of "a" in "panamabananas$":
pa3 na2 ma1 ba4 na5 na6 s$
EXERCISE BREAK: Where are the three instances of "n" from FirstColumn (i.e.,
"n1 ", "n2 ", and "n3 ") located in "panamabananas$"?
To locate "a1 " in LastColumn, we need to cyclically rotate the second row of the ma-
trix M ("a1 bananas$panam"), which results in "bananas$panama1 ". This rotation corre-
sponds to the eighth row of M ("panamabananas$"):
b a n a n a s $ p a n a m a1
m a b a n a n a s $ p a n a
n a m a b a n a n a s $ p a
n a n a s $ p a n a m a b a
n a s $ p a n a m a b a n a
s $ p a n a m a b a n a n a
142
EXERCISE BREAK: Where are the other five instances of "a" located in LastCol-
umn?
You hopefully saw that LastColumn can be recorded as "smnpbnna1 a2 a3 a4 a5 $a6 ", as
shown in Figure 9.11. Note that the six instances of "a" appear in exactly the same order
in FirstColumn and LastColumn. This observation is not a fluke. On the contrary, it is a
principle that holds for any string Text and any symbol that we choose.
First-Last Property: The k-th occurrence of a symbol in FirstColumn and the k-th occurrence
of this symbol in LastColumn correspond to the same position of this symbol in Text.
m a b a n a n a s $ p a n a2
n a m a b a n a n a s $ p a3
n a n a s $ p a n a m a b a4
n a s $ p a n a m a b a n a5
s $ p a n a m a b a n a n a6
FIGURE 9.11 The six occurrences of "a" occur in the same order in FirstColumn as
they do in LastColumn.
To see why the First-Last Property is true, consider the rows of M ("panamabananas$")
beginning with "a":
These rows are already ordered lexicographically, so if we chop off the "a" from the be-
ginning of each row, then the remaining strings should still be ordered lexicographically:
143
CHAPTER 9
b a n a n a s $ p a n a m
m a b a n a n a s $ p a n
n a m a b a n a n a s $ p
n a n a s $ p a n a m a b
n a s $ p a n a m a b a n
s $ p a n a m a b a n a n
Adding "a" back to the end of each row should not change the lexicographic ordering
of these rows:
m a b a n a n a s $ p a n a2
n a m a b a n a n a s $ p a3
n a n a s $ p a n a m a b a4
n a s $ p a n a m a b a n a5
s $ p a n a m a b a n a n a6
But these are just the rows of M ("panamabananas$") containing "a" in LastColumn! As
a result, the k-th occurrence of "a" in FirstColumn corresponds to the k-th occurrence of
"a" in LastColumn. This argument generalizes for any symbol and any string Text, which
establishes the First-Last property.
Using the First-Last property to invert the Burrows-Wheeler transform
The First-Last Property is interesting, but how can we use it to invert BWT (Text) =
"ard$rcaaaabb"? Recalling Figure 9.10, let’s return to where we were in our attempt
to reconstruct the first row of M (Text) and index the occurrences of each symbol in
FirstColumn and LastColumn:
$1 a ? ? ? ? ? ? ? ? ? ? ? a1
a1 ? ? ? ? ? ? ? ? ? ? ? ? r1
a2 ? ? ? ? ? ? ? ? ? ? ? ? d1
a3 ? ? ? ? ? ? ? ? ? ? ? ? $1
a4 ? ? ? ? ? ? ? ? ? ? ? ? r2
a5 ? ? ? ? ? ? ? ? ? ? ? ? c1
b1 ? ? ? ? ? ? ? ? ? ? ? ? a2
b2 ? ? ? ? ? ? ? ? ? ? ? ? a3
c1 ? ? ? ? ? ? ? ? ? ? ? ? a4
d1 ? ? ? ? ? ? ? ? ? ? ? ? a5
r1 ? ? ? ? ? ? ? ? ? ? ? ? b1
r2 ? ? ? ? ? ? ? ? ? ? ? ? b2
The First-Last Property reveals where "a3 " is hiding in LastColumn:
144
$1 a ? ? ? ? ? ? ? ? ? ? ? a1
a1 ? ? ? ? ? ? ? ? ? ? ? ? r1
a2 ? ? ? ? ? ? ? ? ? ? ? ? d1
a3 ? ? ? ? ? ? ? ? ? ? ? ? $1
a4 ? ? ? ? ? ? ? ? ? ? ? ? r2
a5 ? ? ? ? ? ? ? ? ? ? ? ? c1
b1 ? ? ? ? ? ? ? ? ? ? ? ? a2
b2 ? ? ? ? ? ? ? ? ? ? ? ? a3
c1 ? ? ? ? ? ? ? ? ? ? ? ? a4
d1 ? ? ? ? ? ? ? ? ? ? ? ? a5
r1 ? ? ? ? ? ? ? ? ? ? ? ? b1
r2 ? ? ? ? ? ? ? ? ? ? ? ? b2
Since we know that "a3 " is located at the end of the eighth row, we can wrap around
this row to determine that "b2 " follows "a3 " in Text. Thus, the second symbol of Text is
"b", which we can now add to the first row of M (Text):
$1 a b ? ? ? ? ? ? ? ? ? ? a1
a1 ? ? ? ? ? ? ? ? ? ? ? ? r1
a2 ? ? ? ? ? ? ? ? ? ? ? ? d1
a3 ? ? ? ? ? ? ? ? ? ? ? ? $1
a4 ? ? ? ? ? ? ? ? ? ? ? ? r2
a5 ? ? ? ? ? ? ? ? ? ? ? ? c1
b1 ? ? ? ? ? ? ? ? ? ? ? ? a2
b2 ? ? ? ? ? ? ? ? ? ? ? ? a3
c1 ? ? ? ? ? ? ? ? ? ? ? ? a4
d1 ? ? ? ? ? ? ? ? ? ? ? ? a5
r1 ? ? ? ? ? ? ? ? ? ? ? ? b1
r2 ? ? ? ? ? ? ? ? ? ? ? ? b2
In Figure 9.12, we illustrate repeated applications of the First-Last Property to recon-

struct more and more symbols from Text. Presto — the string that we have been trying
to reconstruct is "abracadabra$".
EXERCISE BREAK: Reconstruct the string whose Burrows-Wheeler transform

is "enwvpeoseu$llt".
STOP and Think: Can any string (having a single "$" symbol) be inverted using
the inverse Burrows-Wheeler transform?
You are now ready to implement the inverse of the Burrows-Wheeler transform.
145
CHAPTER 9
$1 a b r ? ? ? ? ? ? ? a1 $1 a b r a ? ? ? ? ? ? a1
a1 ? ? ? ? ? ? ? ? ? ? r1 a1 ? ? ? ? ? ? ? ? ? ? r1
a2 ? ? ? ? ? ? ? ? ? ? d1 a2 ? ? ? ? ? ? ? ? ? ? d1
a3 ? ? ? ? ? ? ? ? ? ? $1 a3 ? ? ? ? ? ? ? ? ? ? $1
a4 ? ? ? ? ? ? ? ? ? ? r2 a4 ? ? ? ? ? ? ? ? ? ? r2
a5 ? ? ? ? ? ? ? ? ? ? c1 a5 ? ? ? ? ? ? ? ? ? ? c1
1 b1 ? ? ? ? ? ? ? ? ? ? a2 2 b1 ? ? ? ? ? ? ? ? ? ? a2
b2 ? ? ? ? ? ? ? ? ? ? a3 b2 ? ? ? ? ? ? ? ? ? ? a3
c1 ? ? ? ? ? ? ? ? ? ? a4 c1 ? ? ? ? ? ? ? ? ? ? a4
d1 ? ? ? ? ? ? ? ? ? ? a5 d1 ? ? ? ? ? ? ? ? ? ? a5
r1 ? ? ? ? ? ? ? ? ? ? b1 r1 ? ? ? ? ? ? ? ? ? ? b1
r2 ? ? ? ? ? ? ? ? ? ? b2 r2 ? ? ? ? ? ? ? ? ? ? b2
$1 a b r a c ? ? ? ? ? a1 $1 a b r a c a ? ? ? ? a1
a1 ? ? ? ? ? ? ? ? ? ? r1 a1 ? ? ? ? ? ? ? ? ? ? r1
a2 ? ? ? ? ? ? ? ? ? ? d1 a2 ? ? ? ? ? ? ? ? ? ? d1
a3 ? ? ? ? ? ? ? ? ? ? $1 a3 ? ? ? ? ? ? ? ? ? ? $1
a4 ? ? ? ? ? ? ? ? ? ? r2 a4 ? ? ? ? ? ? ? ? ? ? r2
a5 ? ? ? ? ? ? ? ? ? ? c1 a5 ? ? ? ? ? ? ? ? ? ? c1
3 b1 ? ? ? ? ? ? ? ? ? ? a2 4 b1 ? ? ? ? ? ? ? ? ? ? a2
b2 ? ? ? ? ? ? ? ? ? ? a3 b2 ? ? ? ? ? ? ? ? ? ? a3
c1 ? ? ? ? ? ? ? ? ? ? a4 c1 ? ? ? ? ? ? ? ? ? ? a4
d1 ? ? ? ? ? ? ? ? ? ? a5 d1 ? ? ? ? ? ? ? ? ? ? a5
r1 ? ? ? ? ? ? ? ? ? ? b1 r1 ? ? ? ? ? ? ? ? ? ? b1
r2 ? ? ? ? ? ? ? ? ? ? b2 r2 ? ? ? ? ? ? ? ? ? ? b2
$1 a b r a c a d ? ? ? a1 $1 a b r a c a d a ? ? a1
a1 ? ? ? ? ? ? ? ? ? ? r1 a1 ? ? ? ? ? ? ? ? ? ? r1
a2 ? ? ? ? ? ? ? ? ? ? d1 a2 ? ? ? ? ? ? ? ? ? ? d1
a3 ? ? ? ? ? ? ? ? ? ? $1 a3 ? ? ? ? ? ? ? ? ? ? $1
a4 ? ? ? ? ? ? ? ? ? ? r2 a4 ? ? ? ? ? ? ? ? ? ? r2
a5 ? ? ? ? ? ? ? ? ? ? c1 a5 ? ? ? ? ? ? ? ? ? ? c1
5 b1 ? ? ? ? ? ? ? ? ? ? a2 6 b1 ? ? ? ? ? ? ? ? ? ? a2
b2 ? ? ? ? ? ? ? ? ? ? a3 b2 ? ? ? ? ? ? ? ? ? ? a3
c1 ? ? ? ? ? ? ? ? ? ? a4 c1 ? ? ? ? ? ? ? ? ? ? a4
d1 ? ? ? ? ? ? ? ? ? ? a5 d1 ? ? ? ? ? ? ? ? ? ? a5
r1 ? ? ? ? ? ? ? ? ? ? b1 r1 ? ? ? ? ? ? ? ? ? ? b1
r2 ? ? ? ? ? ? ? ? ? ? b2 r2 ? ? ? ? ? ? ? ? ? ? b2
$1 a b r a c a d a b ? a1 $1 a b r a c a d a b r a1
a1 ? ? ? ? ? ? ? ? ? ? r1 a1 ? ? ? ? ? ? ? ? ? ? r1
a2 ? ? ? ? ? ? ? ? ? ? d1 a2 ? ? ? ? ? ? ? ? ? ? d1
a3 ? ? ? ? ? ? ? ? ? ? $1 a3 ? ? ? ? ? ? ? ? ? ? $1
a4 ? ? ? ? ? ? ? ? ? ? r2 a4 ? ? ? ? ? ? ? ? ? ? r2
a5 ? ? ? ? ? ? ? ? ? ? c1 a5 ? ? ? ? ? ? ? ? ? ? c1
7 b1 ? ? ? ? ? ? ? ? ? ? a2 8 b1 ? ? ? ? ? ? ? ? ? ? a2
b2 ? ? ? ? ? ? ? ? ? ? a3 b2 ? ? ? ? ? ? ? ? ? ? a3
c1 ? ? ? ? ? ? ? ? ? ? a4 c1 ? ? ? ? ? ? ? ? ? ? a4
d1 ? ? ? ? ? ? ? ? ? ? a5 d1 ? ? ? ? ? ? ? ? ? ? a5
r1 ? ? ? ? ? ? ? ? ? ? b1 r1 ? ? ? ? ? ? ? ? ? ? b1
r2 ? ? ? ? ? ? ? ? ? ? b2 r2 ? ? ? ? ? ? ? ? ? ? b2
FIGURE 9.12 Repeated applications of the First-Last Property reconstruct the string
"abracadabra$" from its Burrows-Wheeler transform "ard$rcaaaabb".
146
Inverse Burrows-Wheeler Transform Problem:

Reconstruct a string from its Burrows-Wheeler transform.
9J
Input: A string Transform (with a single "$" symbol).
Output: The string Text such that BWT (Text) = Transform.
Pattern Matching with the Burrows-Wheeler Transform
A first attempt at Burrows-Wheeler pattern matching
The Burrows-Wheeler transform may be fascinating, but how could it possibly help us
decrease the memory required for pattern matching? The idea motivating a Burrows-
Wheeler-based approach to pattern matching relies on the observation that each row of
M (Text) begins with a different suffix of Text. Since these suffixes are already ordered
lexicographically, as we already noted when pattern matching with the suffix array, any
matches of Pattern in Text will appear at the beginning of consecutive rows of M (Text),
as shown in Figure 9.13.
M(Text) SUFFIXARRAY(Text)
$ p a n a m a b a n a n a s 13
a b a n a n a s $ p a n a m 5
a m a b a n a n a s $ p a n 3
a n a m a b a n a n a s $ p 1
a n a n a s $ p a n a m a b 7
a n a s $ p a n a m a b a n 9
a s $ p a n a m a b a n a n 11
b a n a n a s $ p a n a m a 6
m a b a n a n a s $ p a n a 4
n a m a b a n a n a s $ p a 2
n a n a s $ p a n a m a b a 8
n a s $ p a n a m a b a n a 10
p a n a m a b a n a n a s $ 0
s $ p a n a m a b a n a n a 12
FIGURE 9.13 (Left) Because the rows of M(Text) are ordered lexicographically, suffixes
beginning with the same string ("ana") appear in consecutive rows of the matrix. (Right)
The suffix array records the starting position of each suffix in Text and immediately tells
us the locations of "ana".
147
CHAPTER 9
We now have the outline of a method to match Pattern to Text. Construct M (Text),
and then identify rows beginning with the first symbol of Pattern. Among these rows,
determine which ones have a second element matching the second symbol of Pattern.
Continue this process until we find which rows of M (Text) begin with Pattern.
STOP and Think: What is wrong with this approach?
Moving backward through a pattern
The issue with this proposed method for pattern matching is that we cannot afford
storing the entire matrix M (Text), which has |Text|2 entries. In an effort to reduce mem-
ory requirements, let’s forbid ourselves from accessing any information in M (Text)
other than FirstColumn and LastColumn. Using these two columns, we will try to
match Pattern to Text by moving backward through Pattern. For example, if we want to
match Pattern = "ana" to Text = "panamabananas$", then we will first identify rows of
M (Text) beginning with "a", the last letter of "ana":
$1 p a n a m a b a n a n a s1
a1 b a n a n a s $ p a n a m1
a2 m a b a n a n a s $ p a n1
a3 n a m a b a n a n a s $ p1
a4 n a n a s $ p a n a m a b1
a5 n a s $ p a n a m a b a n2
a6 s $ p a n a m a b a n a n3
b1 a n a n a s $ p a n a m a1
m1 a b a n a n a s $ p a n a2
n1 a m a b a n a n a s $ p a3
n2 a n a s $ p a n a m a b a4
n3 a s $ p a n a m a b a n a5
p1 a n a m a b a n a n a s $1
s1 $ p a n a m a b a n a n a6
As we are moving backward through "ana", we will next look for rows of M (Text)
beginning with "na". To do this without knowing the entire matrix M (Text), we again
use the fact that a symbol in LastColumn must precede the symbol of Text found in
the same row in FirstColumn. Thus, we only need to identify those rows of M (Text)
beginning with "a" and ending with "n":
148
The First-Last Property tells us where to find the three highlighted "n" in FirstColumn,
as shown below. All three rows end with "a", yielding three total occurrences of "ana"
in Text.
The highlighted occurrences of "a" in LastColumn correspond to the third, fourth, and
fifth occurrences of "a" in this column, and the First-Last Property tells us that they
should correspond to the third, fourth, and fifth occurrences of "a" in FirstColumn as
well, which identifies the three matches of "ana":
149
CHAPTER 9
EXERCISE BREAK: Match Pattern = "banana" to Text = "panamabananas$" by

walking backward through Pattern using the Burrows-Wheeler transform of Text.
The Last-to-First mapping
We now know how to use BWT (Text) to find all matches of Pattern in Text by walking
backward through Pattern. However, every time we walk backward, we need to
keep track of the rows of M (Text) where the matches of a suffix of Pattern are hiding.
Fortunately, we know that at each step, the rows of M (Text) that match a suffix of
Pattern clump together in consecutive rows of M (Text). This means that the collection
of all matching rows is revealed by only two pointers, top and bottom: top holds the
index of the first row of M (Text) that matches the current suffix of Pattern, and bottom
holds the index of the last row of M (Text) that matches this suffix. Figure 9.14 shows
the process of updating pointers; after walking backward through Pattern = "ana",
we have that top = 3 and bottom = 5. After traversing Pattern, we can compute the
total number of matches of Pattern in Text by calculating bottom top + 1 (e.g., there are
5 3 + 1 = 3 matches of "ana" in "panamabananas$").
Let’s concentrate on how pointers are updated from one stage to the next. Consider
the transition from the second to the third panel in Figure 9.14; how did we know to
update the pointers (top = 1, bottom = 6) into (top = 9, bottom = 11)? We are looking for
the first and last occurrence of "n" in the range of positions from top = 1 to bottom = 6
in LastColumn. The first occurrence of "n" in this range is "n1 " (in position 2) and the
last is "n3 " (position 6).
In order to update the top and bottom pointers, we need to determine where "n1 "
and "n3 " occur in FirstColumn. The Last-to-First mapping, denoted L AST T O F IRST (i ),
150
answers the following question: given a symbol at position i in LastColumn, what is its
position in FirstColumn? 9K
For our ongoing example, L AST T O F IRST (2) = 9, since the symbol at position 2 of
LastColumn ("n1 ") occurs at position 9 in FirstColumn, as shown in Figure 9.15. Similarly,
L AST T O F IRST (6) = 11, since the symbol at position 6 of LastColumn ("n3 ") occurs at
position 11 in FirstColumn. Therefore, with the help of the Last-to-First mapping, we
can quickly update the pointers (top = 1, bottom = 6) into (top = 9, bottom = 11).
We are now ready to describe BWM ATCHING, an algorithm that counts the total
number of matches of Pattern in Text, where the only information that we are given is
FirstColumn and LastColumn in addition to the Last-to-First mapping. The pointers top
and bottom are updated by the green lines in the following pseudocode.
BWMATCHING(FirstColumn, LastColumn, Pattern, L ASTTOFIRST)

top 0
bottom |LastColumn| 1
while top  bottom
if Pattern is nonempty
symbol last letter in Pattern
remove last letter from Pattern
if positions from top to bottom in LastColumn contain symbol
topIndex first position of symbol among positions from top to bottom 9L
in LastColumn
bottomIndex last position of symbol among positions from top to
bottom in LastColumn
top L ASTTOFIRST (topIndex)
bottom L ASTTOFIRST (bottomIndex)
else
return 0
else
return bottom top + 1
151
CHAPTER 9
ana ana ana ana

0
$1panamabananas1 1 $1panamabananas1 $1panamabananas1 $1panamabananas1
top
a1bananas$panam1 a1bananas$panam1 a1bananas$panam1 a1bananas$panam1
a2mabananas$pan1 a2mabananas$pan1 a2mabananas$pan1 a2mabananas$pan1
3
a3namabananas$p1 a3namabananas$p1 a3namabananas$p1 a3namabananas$p1
a4nanas$panamab1 a4nanas$panamab1 a4nanas$panamab1 a4nanas$panamab1
a5nas$panamaban2 a5nas$panamaban2 a5nas$panamaban2
5 a5nas$panamaban2
6
a6s$panamabanan3 a6s$panamabanan3 a6s$panamabanan3 a6s$panamabanan3
b1ananas$panama1 b1ananas$panama1 b1ananas$panama1 b1ananas$panama1
m1abananas$pana2 m1abananas$pana2 9 m1abananas$pana2 m1abananas$pana2
n1amabananas$pa3 n1amabananas$pa3 n1amabananas$pa3 n1amabananas$pa3
n2anas$panamaba4 n2anas$panamaba4 n2anas$panamaba4 n2anas$panamaba4
11
n3as$panamabana5 n3as$panamabana5 n3as$panamabana5 n3as$panamabana5
p1anamabananas$1 p1anamabananas$1 p1anamabananas$1 p1anamabananas$1
13
s1$panamabanana6 s1$panamabanana6 s1$panamabanana6 s1$panamabanana6
bottom
FIGURE 9.14 The pointers top and bottom hold the indices of the first and last rows
of M(Text) matching the current suffix of Pattern = "ana". The above diagram shows
how these pointers are updated when walking backwards through "ana" and looking
for substring matches in "panamabananas$".
i FirstColumn LastColumn L ASTTOFIRST(i) COUNT

$ a b m n p s
0 $1 s1 13 0 0 0 0 0 0 0
1 a1 m1 8 0 0 0 0 0 0 1
2 a2 n1 9 0 0 0 1 0 0 1
3 a3 p1 12 0 0 0 1 1 0 1
4 a4 b1 7 0 0 0 1 1 1 1
5 a5 n2 10 0 0 1 1 1 1 1
6 a6 n3 11 0 0 1 1 2 1 1
7 b1 a1 1 0 0 1 1 3 1 1
8 m1 a2 2 0 1 1 1 3 1 1
9 n1 a3 3 0 2 1 1 3 1 1
10 n2 a4 4 0 3 1 1 3 1 1
11 n3 a5 5 0 4 1 1 3 1 1
12 p1 $1 0 0 5 1 1 3 1 1
13 s1 a6 6 1 5 1 1 3 1 1
1 6 1 1 3 1 1
FIGURE 9.15 The Last-to-First mapping and count array. Precomputing the count array
prevents time-consuming updates of the top and bottom pointers in BWMATCHING.
152
Speeding Up Burrows-Wheeler Pattern Matching
Substituting the Last-to-First mapping with count arrays
If you implemented BWM ATCHING in the previous section, you probably found this
algorithm to be slow. The reason for its sluggishness is that updating the pointers top
and bottom is time-intensive, since it requires examining every symbol in LastColumn
between top and bottom at each step. To improve BWM ATCHING, we introduce a func-
tion C OUNTsymbol (i, LastColumn), which returns the number of occurrences of symbol in
the first i positions of LastColumn. For example, C OUNT"n" (10, "smnpbnnaaaaa$a") = 3,
and C OUNT"a" (4, "smnpbnnaaaaa$a") = 0. In Figure 9.15, we show arrays holding
C OUNTsymbol (i, "smnpbnnaaaaa$a") for every symbol occurring in "panamabananas$".
EXERCISE BREAK: Compute the arrays C OUNT for BWT ("abracadabra$").
We will say that the k-th occurrence of symbol in a column of a matrix has rank k in
this column. For Text = "panamabananas$", note that the first and last occurrences of
symbol in the range of positions from top to bottom in LastColumn have respective ranks
C OUNTsymbol (top, LastColumn) + 1
and
C OUNTsymbol (bottom + 1, LastColumn).
As illustrated in Figure 9.15, when top = 1, bottom = 6, and symbol = "n",

C OUNT"n" (top, LastColumn) + 1 = 1
C OUNT"n" (bottom + 1, LastColumn) = 3
The occurrences of "n" having ranks 1 and 3 are located at positions 2 and 6 of
LastColumn, implying that we should update top to L AST T O F IRST (2) = 9 and bot-
tom to L AST T O F IRST (6) = 11. Thus, the four green lines in the pseudocode for
BWM ATCHING can be rewritten as follows.
topIndex position of symbol with rank COUNTsymbol (top, LastColumn) + 1

in LastColumn
bottomIndex position of symbol with rank COUNTsymbol (bottom + 1, LastColumn)
in LastColumn
top L ASTTOFIRST (topIndex)
bottom L ASTTOFIRST (bottomIndex)
153
CHAPTER 9
By eliminating the variables topIndex and bottomIndex, we can reduce these four lines of
pseudocode to only two lines:
top L ASTTOFIRST (position of symbol with rank COUNTsymbol (top, LastColumn) +

1 in LastColumn)
bottom L ASTTOFIRST (position of symbol with rank COUNTsymbol (bottom + 1,
LastColumn) in LastColumn)
Note that these two lines of pseudocode merely compute the position of symbol with
rank i in FirstColumn from its positions in LastColumn. This task can be compactly
described without the Last-to-First mapping by the following two lines:
top position of symbol with rank COUNTsymbol (top, LastColumn) + 1 in

FirstColumn
bottom position of symbol with rank COUNTsymbol (bottom + 1, LastColumn) in
FirstColumn
For top = 1, bottom = 6, and symbol = "n", the occurrences of "n" having ranks
C OUNT"n" (top, LastColumn) + 1 = 1 and C OUNT"a" (bottom + 1, LastColumn) = 3 are
located in positions 9 and 11 of FirstColumn, respectively.
STOP and Think: Do we need to store all of FirstColumn in memory in order to

execute the preceding two lines of pseudocode?
Getting rid of the first column of the Burrows-Wheeler matrix
BWM ATCHING requires us to store FirstColumn, LastColumn, and L AST T O F IRST. We

can reduce the amount of memory needed to store the information contained in First-
Column by defining F IRST O CCURRENCE (symbol) as the first position of symbol in First-
Column. If Text = "panamabananas$", then FirstColumn is "$aaaaaabmnnnps", and the
array holding all values of F IRST O CCURRENCE is [0, 1, 7, 8, 9, 11, 12], as shown in Fig-
ure 9.16. For DNA strings of any length, the array F IRST O CCURRENCE contains only
five elements. The previous two lines of code can now be rewritten as follows:
top FIRSTOCCURRENCE (symbol) + COUNTsymbol (top, LastColumn)

bottom FIRSTOCCURRENCE (symbol) + COUNTsymbol (bottom+ 1, LastColumn) 1
154
i FirstColumn FIRSTOCCURRENCE
0 $1 0
1 a1 1
2 a2
3 a3
4 a4
5 a5
6 a6
7 b1 7
8 m1 8
9 n1 9
10 n2
11 n3
12 p1 12
13 s1 13
FIGURE 9.16 The array FIRSTOCCURRENCE has just seven elements, which is equal to
the number of distinct symbols in "panamabananas$".
When top = 1, bottom = 6, and symbol = "n", we have that
F IRST O CCURRENCE ("n") = 9

C OUNT"n" (top, LastColumn) = 0
C OUNT"n" (bottom + 1, LastColumn) = 3
Recalling Figure 9.14, this again implies that (top = 1, bottom = 6) will be updated as
top = 9 + 0 = 9
bottom = 9 + 3 1 = 11
In the process of simplifying the green lines of pseudocode from BWM ATCHING, we
have also substituted FirstColumn by F IRST O CCURRENCE and L AST T O F IRST by C OUNT,
resulting in a more efficient algorithm called B ETTER BWM ATCHING, shown below.
You may be wondering why we call this algorithm “better”, since on the one hand,
if you have to compute the C OUNT arrays as you go, then you will not obtain a runtime
speedup. On the other hand, if you have to precompute these arrays, then you will
need to store them in memory, which is space-intensive. Hold onto this thought.
155
CHAPTER 9
BETTERBWMATCHING(FIRSTOCCURRENCE, LastColumn, Pattern, COUNT)

top 0
bottom |LastColumn| 1
while top  bottom
if Pattern is nonempty
symbol last letter in Pattern
9M remove last letter from Pattern
top FIRSTOCCURRENCE (symbol) + COUNTsymbol (top, LastColumn)
bottom FIRSTOCCURRENCE (symbol) + COUNTsymbol (bottom + 1,
LastColumn) 1
else
return bottom top + 1
return
Where are the Matched Patterns?
We hope that you have noticed a limitation of B ETTER BWM ATCHING — even though
this algorithm counts the number of occurrences of Pattern in Text, it does not tell
us where these occurrences are located in Text! To locate pattern matches identified
by the algorithm, we can once again use the suffix array, as shown in Figure 9.13
(right). In this figure, the suffix array immediately finds the three matches of "ana" in
"panamabananas$".
The suffix array makes our job easy, but recall that our original motivation for using
the Burrows-Wheeler transform was to reduce the amount of memory used by the suffix
array for pattern matching. If we add the suffix array to Burrows-Wheeler-based pattern
matching, then we are right back where we started!
The memory-saving device that we will employ is inelegant but useful. We will
build a partial suffix array of Text, denoted S UFFIX A RRAY K (Text), which only contains
values that are multiples of some positive integer K (Figure 9.17). In real applications,
partial suffix arrays are often constructed for K = 100, thus reducing memory usage
by a factor of 100 compared to a full suffix array. See CHARGING STATION: Partial
PAGE Suffix Array Construction for more details.
169
156
Partial
panamabananas$ panamabananas$ panamabananas$ Array
$1panamabananas1 $1panamabananas1 $1panamabananas1 13

a1bananas$panam1 a1bananas$panam1 a1bananas$panam1 5
a2mabananas$pan1 a2mabananas$pan1 a2mabananas$pan1 3
a3namabananas$p1 a3namabananas$p1 a3namabananas$p1 1
a4nanas$panamab1 a4nanas$panamab1 a4nanas$panamab1 7
a5nas$panamaban2 a5nas$panamaban2 a5nas$panamaban2 9
a6s$panamabanan3 a6s$panamabanan3 a6s$panamabanan3 11
b1ananas$panama1 b1ananas$panama1 b1ananas$panama1 6
m1abananas$pana2 m1abananas$pana2 m1abananas$pana2 4
n1amabananas$pa3 n1amabananas$pa3 n1amabananas$pa3 2
n2anas$panamaba4 n2anas$panamaba4 n2anas$panamaba4 8
n3as$panamabana5 n3as$panamabana5 n3as$panamabana5 10
p1anamabananas$1 p1anamabananas$1 p1anamabananas$1 0
s1$panamabanana6 s1$panamabanana6 s1$panamabanana6 12
FIGURE 9.17 One of the matches of "ana" in "panamabananas$" is highlighted in the

matrix on the right. By walking backward, we find that "ana" is preceded by "b1 ",
which in turn is preceded by "a1 ". The partial suffix array above, generated for K = 5,
indicates that "a1 " occurs at position 5 of "panamabananas". Since it took us two
steps to walk backward to "a1 ", we conclude that this occurrence of "ana" begins at
position 5 + 2 = 7.
Burrows and Wheeler Set Up Checkpoints
We will now discuss how to improve B ETTER BWM ATCHING by resolving the trade-off
between precomputing the values of C OUNTsymbol (i, LastColumn) (requiring substantial
memory) and computing these values as we go (requiring substantial runtime).
The balance that we strike is similar to the one used for the partial suffix array.
Rather than storing C OUNTsymbol (i, LastColumn) for all positions i, we will only store
the C OUNT arrays when i is divisible by C, where C is a constant; these arrays are
called checkpoint arrays (Figure 9.18). When C is large (C is typically equal to 100 in
practice) and the alphabet is small (e.g., four nucleotides), checkpoint arrays require
only a fraction of the memory used by BWT (Text).
What about runtime? Using checkpoint arrays, we can compute the top and bottom
pointers in a constant number of steps (i.e., fewer than C). Since each string Pattern
requires at most |Pattern| pointer updates, the modified B ETTER BWM ATCHING algo-
rithm now requires O(|Patterns|) runtime, which is the same as using a trie or suffix
array.
Furthermore, we now only have to store the following data in memory: BWT (Text),
F IRST O CCURRENCE, the partial suffix array, and the checkpoint arrays. Storing this
data requires memory approximately equal to 1.5 · |Text|. Thus, we have finally knocked
157
CHAPTER 9
i LastColumn COUNT
$ a b m n p s
0 s1 0 0 0 0 0 0 0
1 m1 0 0 0 0 0 0 1
2 n1 0 0 0 1 0 0 1
3 p1 0 0 0 1 1 0 1
4 b1 0 0 0 1 1 1 1
5 n2 0 0 1 1 1 1 1
6 n3 0 0 1 1 2 1 1
7 a1 0 0 1 1 3 1 1
8 a2 0 1 1 1 3 1 1
9 a3 0 2 1 1 3 1 1
10 a4 0 3 1 1 3 1 1
11 a5 0 4 1 1 3 1 1
12 $1 0 5 1 1 3 1 1
13 a6 1 5 1 1 3 1 1
1 6 1 1 3 1 1
FIGURE 9.18 The COUNT checkpoint arrays for Text = "panamabananas$" and C = 5
are highlighted in bold. If we want to compute COUNT"a" (13, "smnpbnnaaaaa$a"),
then the checkpoint array at position 10 tells us that there are 3 occurrences of
"a" before position 10 of "smnpbnnaaaaa$a". We then check whether "a" is
present at position 10 (yes), 11 (yes), and 12 (no) of LastColumn to conclude that
COUNT"a" (13, "smnpbnnaaaaa$a") = 3 + 2 = 5.
down the memory required for solving the Multiple Pattern Matching Problem for
9N
millions of sequencing reads into a workable range.
As the plummeting cost of DNA sequencing has made headlines, it is easy to fail to
appreciate the progress that has been made in the computational side of read mapping.
So, before continuing into approximate pattern matching, we would like to pause and
reflect on just how far the algorithms for read mapping – and the hardware running
them – have progressed. In 1975, the state-of-the-art Aho-Corasick algorithm was touted
as requiring 15 minutes to map just 24 English words to a dictionary. Just a generation
later, when Burrows-Wheeler based approaches were introduced into read mapping,
the same 15 minutes was used to map almost ten million reads to a reference genome.
Having seen how far we have come in the last four decades, we can only imagine where
it will take us in the future.
158
Epilogue: Mismatch-Tolerant Read Mapping
Reducing approximate pattern matching to exact pattern matching
In this section, we will return to the goal of identifying SNPs in an individual genome
when compared against the reference genome. To do so, we need to generalize the
Approximate Pattern Matching Problem from Chapter 1 to the case of multiple patterns.
Multiple Approximate Pattern Matching Problem:

Find all approximate occurrences of a collection of patterns in a text.
Input: A string Text, a collection of shorter strings Patterns, and an integer d.

Output: All starting positions in Text where a string from Patterns appears
as a substring with at most d mismatches.
We begin with the simple observation that if Pattern occurs in Text with a single mis-
match, then we can divide Pattern into two halves, one of which occurs exactly in Text,
as illustrated below.
Pattern acttggct
Text ...ggcacactaggctcc...
Thus, we can find whether Pattern occurs with one mismatch in Text by dividing Pattern
into two halves and then searching for exact matches of these shorter strings. If we find
a match with one of these halves of Pattern, we then check if the entire string Pattern
occurs with a single mutation.
This method can easily be generalized for approximate pattern matching with
d > 1 mismatches; if Pattern approximately matches a substring of Text with at most d
mismatches, then Pattern and Text must share at least one k-mer for a sufficiently large
value of k. For example, if we are looking for a pattern of length 20 with at most d = 3
mismatches, then we can divide this pattern into four parts of length 20 (3 + 1) = 5
and search for exact matches of these shorter substrings:
Pattern acttaggctcgggataatcc
Text ...actaagtctcgggataagcc...
159
CHAPTER 9
This observation is helpful because it reduces approximate pattern matching to the

exact matching of shorter patterns, which allows us to use fast algorithms that are
designed for exact pattern matching but are not applicable to approximate pattern
matching, such as approaches based on suffix trees and suffix arrays.
STOP and Think: If Pattern has length 23 and appears in Text with 3 mismatches,
can we conclude that Pattern shares a 6-mer with Text? Can we conclude that it
shares a 5-mer with Text?
The question remains how to find the maximum value of k that guarantees that any
Pattern of length n with d mismatches in Text will have a k-mer substring exactly
matching Text.
Theorem. If two strings of length n match with at most d mismatches, then they must share a
k-mer of length k = b n (d + 1) c.
Proof. Divide the first string into d + 1 substrings, where the first d substrings have
exactly k symbols and the final substring has at least k symbols. For the case d = 3, here
is a subdivision of a 23-nucleotide string (n = 23) into d + 1 = 4 substrings, where the
first 3 substrings have k = b n (d + 1) c = b23/4c = 5 symbols and the last substring
has 8 symbols.
acttaggctcgggataatccgga
If you distribute d mismatches among the positions of the string, then the mismatches
may affect at most d substrings, which leaves at least one substring (of length at least k)
unchanged. This substring is shared by both strings.
We now have the outline of an algorithm for matching a string Pattern of length n to
Text with at most d mismatches. We first divide Pattern into d + 1 segments of length
k = b n (d + 1) c, called seeds. After finding which seeds match Text exactly (seed
detection), we attempt to extend seeds in both directions in order to verify whether
Pattern occurs in Text with at most d mismatches (seed extension).
BLAST: Comparing a sequence against a database
Using shared k-mers to find similarities between biological sequences has some dis-
advantages. For example, two proteins may have similar functions but not share any
k-mers, even for small values of k.
160
The Basic Local Alignment Search Tool (BLAST) is a heuristic that can find simi-
larities between proteins, even if all amino acids in one protein have mutated compared
to the other protein. BLAST is so fast that it is often used to query a protein against all
other known proteins in protein databases. The paper introducing BLAST was released
in 1990, and with over 40,000 citations, it has become one of the most cited scientific
papers ever published.
To see how BLAST works, say we are given an integer k and strings x = x1 . . . xn
and y = y1 . . . ym to compare. We define a segment pair of x and y as a pair formed by
a k-mer from x and a k-mer from y. The score of the segment pair corresponding to the
k-mers starting at position i in x and position j in y is
k 1
Â S CORE(xi+t , y j+t ) ,
t =0
where S CORE ( xi+t , y j+t ) is determined by a scoring matrix such as the PAM scoring
matrix that we introduced in Chapter 5. A locally maximal segment pair is a segment
pair whose score cannot be increased by extending or shortening both strings in the
segment pair. BLAST attempts to find not the highest-scoring segment pair of x and y,
but rather all locally maximal segment pairs in these strings with scores above some
threshold.
The key ingredient of BLAST is to first quickly find all k-mers that have scores above
a given threshold when scored against some k-mer in the query string. If the score
threshold is high, then the set of all resulting segment pairs formed between the query
string and strings in the database is not too large. In this case, the database can be
searched for exact occurrences of these high-scoring k-mers from this set, producing an
initial set of seeds. This is an instance of the Multiple Pattern Matching Problem, which
we have learned how to solve quickly.
EXERCISE BREAK: Given the PAM250 scoring matrix, only five 3-mers score
higher than 23 against CFC: CIC, CLC, CMC, CWC, and CYC. How many 3-mers
score higher than 20 against CFC?
After finding seeds, BLAST attempts to extend these seeds (allowing for insertions and
deletions) in order to obtain locally maximal segment pairs.
EXERCISE BREAK: Given the PAM250 scoring matrix, an amino acid k-mer
Peptide, and a threshold q, develop an efficient algorithm for finding the exact
number of k-mers scoring more than q against Peptide.
161
CHAPTER 9
Approximate pattern matching with the Burrows-Wheeler transform
To extend the Burrows-Wheeler approach to approximate pattern matching, we will not

stop when we encounter a mismatch. On the contrary, we will proceed onward until
we either find an approximate match or exceed the limit of d mismatches.
Figure 9.19 illustrates the search for "asa" in "panamabananas$" with at most 1
mismatch. Let’s first proceed as in the case of exact pattern matching, threading "asa"
backwards using the Burrows-Wheeler transform. After finding six occurrences of "a",
we identify six inexact occurrences of "sa": "pa", "ma", "ba", and three occurrences of
"na". We note that these three strings have accumulated a mismatch and then continue
with all six strings.
In the next step, five of the inexact occurrences of "sa" can be extended into inexact
occurrences of "asa" with only a single mismatch: "ama", "aba", and three occurrences
of "ana". We fail to extend "pa", which we eliminate from consideration.
In practice, this heuristic faces complications. We do not want to start allowing
mismatched strings at the early stages of B ETTER BWM ATCHING, or else we will have
to consider too many frivolous candidate strings. We may therefore require that a suffix
of Pattern of some threshold length matches Text exactly. Moreover, the method becomes
time-intensive when using large values of d, as we must explore many inexact matches.
Practical applications often limit the value of d to at most 3.
You should now be ready to design your own approach to solve the Multiple
9O Approximate Pattern Matching Problem and use this solution to map real sequencing
reads.
CHALLENGE PROBLEM: Given the Burrows-Wheeler transform and a partial

suffix array of the bacterial genome Mycoplasma pneumoniae along with a collection
of reads, find all reads occurring in the genome with at most one mismatch.
162
asa asa
$1panamabananas1 top = 1 $1panamabananas1
a1bananas$panam1 a1bananas$panam1
a2mabananas$pan1 bottom = 1 a2mabananas$pan1
a3namabananas$p1 a3namabananas$p1
a4nanas$panamab1 a4nanas$panamab1
a5nas$panamaban2 a5nas$panamaban2
top = 7 a6s$panamabanan3 a6s$panamabanan3
b1ananas$panama1 b1ananas$panama1
bottom = 7 m1abananas$pana2 m1abananas$pana2
n1amabananas$pa3 n1amabananas$pa3
n2anas$panamaba4 n2anas$panamaba4
n3as$panamabana5 n3as$panamabana5
p1anamabananas$1 p1anamabananas$1
s1$panamabanana6 s1$panamabanana6
$1panamabananas1 $1panamabananas1
a1bananas$panam1 top = 2 a1bananas$panam1
a2mabananas$pan1 a2mabananas$pan1
a3namabananas$p1 bottom = 2 a3namabananas$p1
a6s$panamabanan3 a6s$panamabanan3
asa top = 8 b1ananas$panama1 b1ananas$panama1
m1abananas$pana2 m1abananas$pana2
top = 1 $1panamabananas1 bottom = 8 n1amabananas$pa3 n1amabananas$pa3
a1bananas$panam1 n2anas$panamaba4 n2anas$panamaba4
a2mabananas$pan1 n3as$panamabana5 n3as$panamabana5
a3namabananas$p1 p1anamabananas$1 p1anamabananas$1
a4nanas$panamab1 s1$panamabanana6 s1$panamabanana6
a5nas$panamaban2
a6s$panamabanan3
bottom = 6 b1ananas$panama1
m1abananas$pana2
$1panamabananas1 $1panamabananas1
n1amabananas$pa3
n2anas$panamaba4
a2mabananas$pan1 top = 3 a2mabananas$pan1
n3as$panamabana5
p1anamabananas$1
s1$panamabanana6
a6s$panamabanan3 bottom = 5 a6s$panamabanan3
top = 9 m1abananas$pana2 m1abananas$pana2
n3as$panamabana5 n3as$panamabana5
bottom = 11 p1anamabananas$1 p1anamabananas$1
s1$panamabanana6 s1$panamabanana6
$1panamabananas1 2 mismatches $1panamabananas1

a2mabananas$pan1 a2mabananas$pan1
a6s$panamabanan3 a6s$panamabanan3
m1abananas$pana2 m1abananas$pana2
top = 12 n3as$panamabana5 n3as$panamabana5
p1anamabananas$1 p1anamabananas$1
bottom = 12 s1$panamabanana6 s1$panamabanana6
FIGURE 9.19 Using the Burrows-Wheeler transform to find approximate pattern matches
of "asa" in "panamabananas$" with 1 mismatch. (Left) We identify six occurrences
of "a". (Middle) Working backwards, we find four different inexact partial matches:
"ba", "ma", "na" (three occurrences), and "pa". (Right) Additional mismatches are
not allowed in the partial matches that we have found, but these partial matches can
be extended to yield five inexact matches of "asa".
163
CHAPTER 9
Charging Stations
Constructing a suffix tree
To construct a suffix tree, we will first modify the construction of the suffix trie as follows.
Although each edge edge in a suffix trie is labeled by a single symbol S YMBOL (edge), it
is unclear where this symbol came from in Text. We will therefore add another label for
each edge (denoted P OSITION (edge)) referring to the position of this symbol in Text. If
S YMBOL (edge) corresponds to more than one position in Text, then we will assign it its
minimum starting position.
For example, consider the modified suffix trie for Text = "panamabananas$" shown
in Figure 9.20. There are five edges labeled by "m" (colored purple), all of which are
root
$,13 s,12
m,4
a,1 p,0
b,6 n,2
13
b,6 s,12 a,7 a,5 a,3 a,1 $,13
n,2
m,4
12
a,7 a,5 a,3 $,13 n,8 b,6 m,4 s,12 n,2
n,10
11
n,8 b,6 m,4 n, s,12 a,9 a,7 a,5 a,11 $,13 a.3
10
10
a,9 a,7 a,5 a,11 $,13 n,10 n,8 b,6 s,12 m,4
9
n,10 n,8 b,6 s,12 a,11 a,9 a,7 $,13 a,5
8
a,11 a,9 a,7 $,13 s,12 n,10 n,8 b,6
7
s,12 n,10 n,8 $,13 a,11 a,9 a,7
6
$,13 a,11 a,9 s,12 n,10 n,8
5
s,12 n,10 $,13 a,11 a,9
4
$,13 a,11 s,12 n,10
3
s,12 $,13 a,11
2
$,13 s,12
$,13
0
FIGURE 9.20 The modified suffix trie of Text = "panamabananas$". Each edge is
assigned the minimum position to which the edge’s symbol corresponds in Text (shown
in blue).
164
labeled by position 4, because this is the only occurrence of "m" in Text. On the other
hand, the green edge corresponding to "n" is a different story. By following every path
from this edge down to the leaves, we see that it corresponds to occurrences of "n" in
the suffixes "anamabananas$", "ananas$", and "anas$", at positions 2, 8, and 10. As a
result, we assign this edge the minimum of these positions.
The following pseudocode constructs the modified suffix trie of a string Text by
traversing the suffixes of Text from longest to shortest. Given a suffix, it attempts to
spell the suffix by moving downward in the tree, following edge labels as far as possible
until it can go no further. At that point, it adds the rest of the suffix to the trie in the
form of a path to a leaf, along with the position of each symbol in the suffix.
MODIFIEDSUFFIXTRIECONSTRUCTION(Text)
Trie a graph consisting of a single node root
for i 0 to |Text| 1
currentNode root
for j i to |Text| 1
currentSymbol j-th symbol of Text
if there is an outgoing edge from currentNode labeled by currentSymbol
currentNode ending node of this edge
else
add a new node newNode to Trie
add an edge newEdge connecting currentNode to newNode in Trie
SYMBOL (newEdge) currentSymbol
POSITION (newEdge) j
currentNode newNode
if currentNode is a leaf in Trie
assign label i to this leaf
return Trie
We can now transform a modified suffix trie into a suffix tree as follows. Note in
Figure 9.6 (page 131) that each edge edge in S UFFIX T REE ("panamabananas$") is labeled
by a string of symbols, denoted S TRING (edge). As we mentioned in the main text,
storing all these strings is memory-intensive, and so we will instead label edge by two
integers: the starting position of the first occurrence of S TRING (edge) in Text, denoted
P OSITION (edge), and its length, denoted L ENGTH (edge). For the modified suffix tree of
Text = "panamabananas$" shown in Figure 9.21, these two integers are colored blue and
red, respectively. For example, the edge labeled "mabananas$" in Figure 9.6 is labeled
165
CHAPTER 9
by P OSITION (edge) = 4 and L ENGTH (edge) = 10 in Figure 9.21. There are two edges
labeled "na" in Figure 9.6, and both of them are labeled by P OSITION (edge) = 2 and
L ENGTH (edge) = 2 in Figure 9.21.
root
$
a m
13 a
b b n p s
a a a a $
n n n
a a a 12
b n s n n m
a m a $ a a a
n a s s n b
a b 11 $ $ a a
n a m s s n
a n 6 4 a $ $
n b a
s a n
n a a 8 10
$ m s s a
a n
a $ $ a s
s b $
5 $ n
a 7 9 a
n s 0
3 a $
n
a 2
s
$
1
root
13,1
1,1
13
2,2 12,2
6,8 4,10 12
2,2 12,2
4,10 11 10,4 0,14

6,8 12,2
6 4
10,4 8 10
12,2
4,10
5
7 9
3 0
4,10 2
FIGURE 9.21 (Top) The suffix tree of Text = "panamabananas$. (Bottom) The modified
suffix tree of Text. For each edge, the initial position of the substring to which it
corresponds in Text is shown in blue, and the length of this substring is shown in red.
The following pseudocode constructs a suffix tree using the modified suffix trie con-
structed by M ODIFIED S UFFIX T REE C ONSTRUCTION. This algorithm will consolidate
each non-branching path (i.e., a path whose intermediate nodes have indegree and
166
outdegree equal to 1) of the modified suffix trie into a single edge.
SUFFIXTREECONSTRUCTION(Text)
Trie MODIFIEDSUFFIXTRIECONSTRUCTION (Text)
for each non-branching path Path in Trie
substitute Path by a single edge e connecting the first and last nodes of Path
POSITION (e) POSITION (first edge of Path)
LENGTH (e) number of edges of Path
return Trie
Solving the Longest Shared Substring Problem
A naive approach for finding a longest shared substring of strings Text1 and Text2 would
construct one suffix tree for Text1 and another for Text2 . Instead, we will add "#" to the
end of Text1 , add "$" to the end of Text2 , and then construct the single suffix tree for the
concatenation of Text1 and Text2 (Figure 9.22). We color a leaf in this suffix tree blue if it
is labeled by the starting position of a suffix starting in Text1 ; we color a leaf red if it is
labeled by the starting position of a suffix starting in Text2 .
We also color the remaining nodes of the suffix tree blue, red, and purple according
to the following rules:
• a node is colored blue or red if all leaves in its subtree (i.e., the subtree beneath it)
are all blue or all red, respectively;
• a node is colored purple if its subtree contains both blue and red leaves.
We use C OLOR (v) to denote the color of node v.

There are three purple nodes in Figure 9.22 (other than the root), and the strings
spelled from the root to each of these nodes are "a", "ana", and "na". Note that these
three substrings are shared by Text1 = "panama" and Text2 = "bananas". This is no
accident.
EXERCISE BREAK: Prove that a path ending in a purple node in the suffix tree
of Text1 and Text2 spells out a substring shared by Text1 and Text2 .
EXERCISE BREAK: Prove that a path ending in a blue (respectively, red) node
in the suffix tree of Text1 and Text2 spells out a substring that appears in Text1 but
not in Text2 (respectively, Text2 but not in Text1 ).
167
CHAPTER 9
root
$
a
14 m s
b n p
# a a $
b a a
n # n
a b 13
n a a
# n s n a m
a m n
b a a $ a a
n a a n
a # s #
n 12 $ n a b
s b a
a a m s s a
$ n s a $
n 7 $ n
a n $ #
6 a a a
s n m b 9 11 n
$ s s 4 a
a a $ $ a
s # n s
5 $ b a $
8 10 n
a a
3 n 0
a s
n $
a
s 2
$
1
FIGURE 9.22 SUFFIXTREE ("panama#bananas$"), constructed for Text1 = "panama"

and Text2 = "bananas". Leaves corresponding to suffixes starting in "panama#" are
colored blue; leaves corresponding to suffixes starting in "bananas$" are colored red.
Every string of symbols spelled from the root to a purple node corresponds to a substring
shared by Text1 and Text2 .
The previous two exercises imply that in order to find the longest shared substring
between Text1 and Text2 , we need to examine all purple nodes as well as the strings
spelled by paths leading to the purple nodes. A longest such string yields a solution to
the Longest Shared Substring Problem.
T REE C OLORING, which is illustrated in Figure 9.23, colors the nodes of a suffix tree
from the leaves upward. This algorithm assumes that the leaves of the suffix tree have
been labeled "blue" or "red" and all other nodes have been labeled "gray". A node in
a tree is called ripe if it is gray but has no gray children.
TREECOLORING(ColoredTree)
while ColoredTree has ripe nodes
for each ripe node v in ColoredTree
if there exist differently colored children of v
COLOR (v) "purple"
else
COLOR (v) color of all children of v
return ColoredTree
168
1 2
3 4
9P
FIGURE 9.23 An illustration of the steps taken by TREECOLORING for a tree with an
initial coloring of leaves as red or blue (top left).
Partial suffix array construction
To construct the partial suffix array S UFFIX A RRAY K (Text), we first need to construct the
full suffix array and then retain only the elements of this array that are divisible by K,
9Q
along with their indices i. This is illustrated in Figure 9.24 for Text = "panamabananas$"
and K = 5, where S UFFIX A RRAY K (Text) corresponds to the bold elements.
i 0 1 2 3 4 5 6 7 8 9 10 11 12 13
SUFFIXARRAY (Text) 13 5 3 1 7 9 11 6 4 3 8 10 0 12
FIGURE 9.24 Partial suffix array construction.
169
CHAPTER 9
Detours
The reference human genome
When genomes are assembled from DNA taken from a number of donors, the reference
genome represents a mosaic of donor genomes. The existing reference human genome
was derived from thirteen volunteers in the United States; it continues to be improved
by fixing errors and filling in the remaining gaps (there are currently over a hundred).
The reference genome is often used as a template on which new individual genomes
can be rapidly assembled. Comparison of the reference genome and individual human
genomes typically reveals about three million SNPs, and about 0.1% of an individual
human genome cannot be matched to the reference genome at all.
In regions with high diversity, the reference genome may differ significantly from
individual genomes. An example of a high-diversity region of the human genome is
the major histocompatibility complex, a family of genes powering immune systems.
Since these genes have an unusually large number of alternate forms, two individuals
hardly ever have exactly the same set of genes from this complex.
Rearrangements, insertions, and deletions in human genomes
Until recently, biologists focused primarily on SNPs in the human genome, assuming
that rearrangements and indels are relatively rare. In 2005, Evan Eichler surprised biol-
ogists when he found hundreds of rearrangements and indels separating the genomes
of two individuals. This finding was important because rearrangements and indels are
often hallmarks of disease; for example, repeated insertions of the nucleotide triplet
CAG increases the severity of Huntington’s disease.
In 2013, Gerton Lunter revealed the true extent of indels in the human population
by identifying over a million indels in a cohort of over a hundred individuals. Intrigu-
ingly, he found that over half of the indels occurred in just 4% of the genome; in other
words, some regions of the human genome represent “indel hotspots”. As the catalog
of human rearrangements and indels grows, biologists are gaining the ability to iden-
tify frequently mutated genes as well as implicate rearrangements and indels when
diagnosing complex disorders.
The Aho-Corasick algorithm
The Aho-Corasick algorithm for the Multiple Pattern Matching Problem was invented
by Alfred Aho and Margaret Corasick in 1975. The runtime of their algorithm is
O(|Patterns| + |Text| + m), where m is the number of output matches.
170
Imagine sliding the trie in Figure 9.25 against Text = "bantenna". Similarly to
T RIE M ATCHING, the Aho-Corasick algorithm starts at the root and attempts to build a
path spelling a prefix of "bantenna". This attempt fails after three nodes ("bantenna").
In T RIE M ATCHING, we would begin again at the root and attempt to find a match
starting at the second symbol of "bantenna".
root
a p
b n
n a a a
Fail
u re e
a d t dge n b n n
n e a d a
a n n a
s n a n
a a
FIGURE 9.25 The trie from Figure 9.1 with an additional failure edge. The failure edge
allows us to jump from the path beginning with "ban" to the path beginning with "an".
However, we have thrown away some important information: we already matched

the first two symbols of "antenna" when searching through "bantenna". Thus, a more
sensible strategy is to jump directly to the node "an" and then continue downward
in the trie. Indeed, this strategy will eventually match the pattern "antenna". We can
implement this “jumping ahead” strategy by augmenting the tree with a failure edge
connecting node "ban" to node "an" (Figure 9.25). More generally, failure edges are
formed by connecting node v to node w if w is the longest suffix of v that appears in the
trie. The Aho-Corasick algorithm follows failure edges whenever a mismatch is found
in order to avoid going all the way back to the root, thus saving time.
From suffix trees to suffix arrays
We can construct S UFFIX A RRAY (Text) from S UFFIX T REE (Text) by applying a preorder
traversal of S UFFIX T REE (Text) as long as the outgoing edges from every node of the
suffix tree are arranged lexicographically.
171
CHAPTER 9
Given a rooted tree, a node w is called a child of a node v if there is an edge connect-
ing v to w in the tree. The preorder traversal of a tree involves visiting a node of the tree,
starting at the root, and then recursively preorder traversing the subtrees rooted at each
of its children from left to right (Figure 9.26), which is accomplished by the following
pseudocode. By taking the order of leaves visited in a preorder traversal of the suffix
tree, we obtain the suffix array (consult Figure 9.6 and Figure 9.7).
B G
C D H
E F I
FIGURE 9.26 The preorder traversal of the above tree visits its nodes in increasing order
of their labels.
PREORDER(Tree, Node)
visit Node
for each child Node’ of Node from left to right
PREORDER (Tree, Node’)
Conversely, S UFFIX T REE (Text) can be constructed from S UFFIX A RRAY (Text) in linear
time by using the longest common prefix (LCP) array of Text, LCP (Text), which stores
the length of the longest common prefix shared by consecutive lexicographically ordered
suffixes of Text. For example, LCP ("panamabananas$") is (0, 0, 1, 1, 3, 3, 1, 0, 0, 0, 2, 2, 0, 0),
as shown in Figure 9.27.
Suffix Tree Construction from Suffix Array Problem:

Construct a suffix tree from the suffix array and LCP array of a string.
9R
Input: A string Text, its suffix array, and its LCP array.
Output: The suffix tree of Text.
172
LCP Array Sorted Suffixes

0
$
0
abananas$
1 |
amabananas$
1 |
anamabananas$
3 |||
ananas$
3 |||
anas$
1 |
as$
0
bananas$
0
mabananas$
0
namabananas$
2 ||
nanas$
2 ||
nas$
0
panamabananas$
0
s$
FIGURE 9.27 The LCP array of "panamabananas$" is formed by sorting the suffixes
of "panamabananas$" lexicographically and then finding the length of the longest
common prefix shared by consecutive suffixes according to the lexicographic order.
From suffix arrays to suffix trees
Given the suffix array S UFFIX A RRAY and LCP array LCP of a string Text, the suffix tree
S UFFIX T REE (Text) can be constructed in linear time using the algorithm illustrated in
Figure 9.28. After constructing a partial suffix tree for the i lexicographically smallest
suffixes (denoted S UFFIX T REEi (Text)), this algorithm iteratively inserts the (i + 1)-th
suffix into this tree to form S UFFIX T REEi+1 (Text).
We define the descent of a node v in a suffix tree, denoted D ESCENT (v), as the length
of the concatenation of all path labels from the root to this node. We assume that the
descents of all nodes in the growing partial suffix tree have been precomputed.
We start with S UFFIX T REE0 (Text), which we define as the tree consisting only of
the root. To insert the (i + 1)-th suffix (corresponding to the element S UFFIX A RRAY (i +
1) in the suffix array of Text) into S UFFIX T REEi (Text), we need to know where the
path representing this suffix “splits” from the already constructed partial suffix tree
173
CHAPTER 9
root root
$ a $ a
13 13
n
m m a
b a b a m
a b a b a
n a n n a b n
a n a a n a a
n a s n a n s
a n $ a n a $
s a s a n
$ s $ s a
$ $ s
$
5 3 5 3 1
root root
$ a $ a
13 13
n n
m a m a
b a m b a
a b a a b m
n a b n n a a
a n a a a n b n
n a n s n a a a
a n a $ a n n s
s a n s a a $
$ s a $ s n
$ s $ a
$ s
5 3 1
$
5 3 1 7
FIGURE 9.28 (Top left) SUFFIXTREE3 (Text) for Text = "panamabananas$". The fourth
lexicographically ordered suffix of Text is "anamabananas$", and we can thread only
its first letter into this suffix tree. Our stopping point is at the purple node, so we
create a new node branching off this node with edge labeled "namabananas$" to
obtain SUFFIXTREE4 (Text) (top right). (Bottom left) The fifth lexicographically or-
dered suffix of Text is "ananas$", the first three symbols of which we can thread
into SUFFIXTREE4 (Text) until we reach a stopping point in the middle of the edge
"namabananas$". (Bottom right) To form SUFFIXTREE5 (Text), we create a new node in
the middle of the purple edge and branch from this node into two edges: one labeled
"mabananas$" in order to retain the suffix "anamabananas$" and another labeled
"nas$" to spell the new suffix "ananas$". In general, the number of red symbols
that we spell down in SUFFIXTREEi (Text) to form SUFFIXTREEi+1 (Text) is given by the
(i + 1)-th entry in the LCP array from Figure 9.27.
S UFFIX T REEi (Text). For example, this split happens at the purple node while insert-
ing "anamabananas$" into S UFFIX T REE3 (Text) (Figure 9.28 (top)) and at the purple
edge labeled "namabananas$" while inserting "ananas$" into S UFFIX T REE4 (Text), thus
breaking this edge into edges labeled "na" and "mabananas$" in S UFFIX T REE5 (Text)
(Figure 9.28 (bottom)).
174
STOP and Think: How would you find the node or edge where the partial suffix
tree splits while constructing the suffix tree from the suffix array and LCP array?
To find the node/edge where the partial suffix tree splits, we will walk up the rightmost
path (i.e., the last added path in the partial suffix tree) in the partial suffix tree beginning
at the previously inserted leaf, labeled S UFFIX A RRAY (i ), to the root. We stop when we
encounter the first node v such that D ESCENT (v)  LCP (i + 1). Afterwards, we have to
consider two cases depending on whether D ESCENT (v) = LCP (i + 1) (the split occurs
at the node v) or D ESCENT (v) < LCP (i + 1) (the split occurs on the edge leading from
v):
• D ESCENT (v) = LCP (i + 1): the concatenation of the labels on the path from
the root to v equals the longest common prefix of suffixes corresponding to
S UFFIX A RRAY (i ) and S UFFIX A RRAY (i + 1). We insert S UFFIX A RRAY (i + 1) as a
new leaf x connected to v, and we label the edge (v, x ) with the suffix of Text
starting at position S UFFIX A RRAY (i + 1) + LCP (i + 1). Thus, the edge label con-
sists of the remaining symbols of the suffix corresponding to S UFFIX A RRAY (i + 1)
that are not already represented by the concatenation of the labels of the path
connecting the root to v. This completes the construction of the partial suffix tree
S UFFIX T REEi+1 (Text) (see Figure 9.28 (top) for an example).
• D ESCENT (v) < LCP (i + 1): the concatenation of the labels on the path from
the root to v has fewer symbols than the longest common prefix of the suffixes
corresponding to S UFFIX A RRAY (i ) and S UFFIX A RRAY (i + 1). The question there-
fore arises of how to recover these missing symbols. We denote the rightmost
edge leading from v in S UFFIX T REEi (Text) as (v, w) and argue that the missing
symbols represent a prefix of this edge’s label. In this case, we split this edge and
construct S UFFIX T REEi+1 (Text) as described below (see Figure 9.28 (bottom) for
an example):
1. Delete the edge (v, w) from S UFFIX T REEi (Text).

2. Add a new internal node y and a new edge (v, y) labeled by a substring
of Text starting at position S UFFIX A RRAY (i + 1) + D ESCENT (v). The new
label is formed by the final LCP (i + 1) D ESCENT (v) symbols of the longest
common prefix of S UFFIX A RRAY (i ) and S UFFIX A RRAY (i + 1). Thus, the con-
catenation of the labels on the path from the root to y is now the longest
common prefix of S UFFIX A RRAY (i ) and S UFFIX A RRAY (i + 1).
175
CHAPTER 9
3. Define D ESCENT (y) as LCP (i + 1).

4. Connect w to the newly created internal node y by an edge (y, w) that is
labeled by a substring of Text starting at position S UFFIX A RRAY (i ) + LCP (i +
1) and ending at position S UFFIX A RRAY (i ) + D ESCENT (w) 1. The new
label consists of the remaining symbols of the deleted edge (v, w) that were
not used as the label of edge (v, y).
5. Add S UFFIX A RRAY (i + 1) as a new leaf x as well as an edge (y, x ) that
is labeled by a suffix of Text beginning at position S UFFIX A RRAY (i + 1) +
LCP (i + 1). The label of this edge consists of the remaining symbols of the
suffix corresponding to S UFFIX A RRAY (i + 1) that are not already represented
by the concatenation of the labels on the path from the root to v.
EXERCISE BREAK: Prove that the running time of this algorithm is O(|Text|).
Binary search
The game show The Price is Right features a timed challenge called the “Clock Game” in
which a contestant makes repeated guesses at the price of an item, with the host telling
the contestant only whether the true price is higher or lower than the most recent guess.
An intelligent strategy for the Clock Game is to pick a sensible range of prices within
which the item’s price must fall, and then guess a price halfway between these two
extremes. If this guess is incorrect, then the contestant has immediately eliminated half
of the set of possible prices. The contestant then makes a guess in the middle of the
remaining possible prices, eliminating half of them again. Iterating this strategy quickly
yields the price of the item.
This strategy for the Clock Game motivates a binary search algorithm finding
the position of an element key within a sorted array A RRAY. This algorithm, called
B INARY S EARCH, is initialized by setting minIndex equal to 0 and maxIndex equal to the
length of A RRAY. It sets midIndex equal to (minIndex + maxIndex)/2 and then checks
to see whether key is greater than or less than A RRAY (midIndex). If key is larger than
this value, then B INARY S EARCH iterates on the subarray of A RRAY from minIndex
to midIndex 1; otherwise, B INARY S EARCH iterates on the subarray of A RRAY from
midIndex + 1 to maxIndex. Iteration eventually identifies the position of key.
For example, if key = 9 and A RRAY = (1, 3, 7, 8, 9, 12, 15), then B INARY S EARCH
would first set minIndex equal to 0, maxIndex equal to 6, and midIndex equal to 3. Because
key is greater than A RRAY (midIndex) = 8, we examine the subarray whose elements
176
are greater than A RRAY (midIndex) by setting minIndex equal to 4, so that midIndex is
recomputed as (4 + 6)/2 = 5. This time, key is smaller than A RRAY (midIndex) = 12, and
so we examine the subarray whose elements are smaller than this value. This subarray
consists of only a single element, which is key.
BINARYSEARCH(ARRAY, key, minIndex, maxIndex)

while maxIndex minIndex
if ARRAY (midIndex) = key
return midIndex
else if ARRAY (midIndex) < key
else
maxIndex midIndex 1
return "key not found"
Bibliography Notes
The Aho-Corasick algorithm was introduced by Aho and Corasick, 1975. Suffix trees
were introduced by Weiner, 1973. Suffix arrays were introduced by Manber and Myers,
1990. The Burrows-Wheeler Transform was introduced by Burrows and Wheeler,
1994. An efficient implementation of the Burrows-Wheeler transform was described
by Ferragina and Manzini, 2000. The genetic cause of Ohdo syndrome was elucidated
by Clayton-Smith et al., 2011. Rearrangements and indels in the human genome were
studied by Tuzun et al., 2005 and Montgomery et al., 2013. BLAST, the dominant
database search tool in molecular biology, was developed by Altschul et al., 1990.
177
W H Y H AV E B I O L O G I S T S S T I L L N O T D E V E L O P E D A N H I V VA C C I N E ?
Classifying the HIV Phenotype
How does HIV evade the human immune system?
In 1984, US Health and Human Services Secretary Margaret Heckler declared that an
HIV vaccine would be available within two years, stating, “Yet another terrible disease
is about to yield to patience, persistence and outright genius.”
In 1997, Bill Clinton established a new research center at the National Institutes of
Health with the goal of developing an HIV vaccine. In his words, “It is no longer a
question of whether we can develop an AIDS vaccine, it is simply a question of when.”
In 2005, Merck began clinical trials of an HIV vaccine but discontinued them two
years later after learning that the vaccine actually increased the risk of HIV infection in
some recipients.
Today, despite enormous investment and ongoing clinical trials, we are still far from
an HIV vaccine, and 35 million people are living with the disease. Scientists have made
great progress in developing a successful antiretroviral therapy, a drug cocktail that
stabilizes an infected patient’s symptoms. However, this therapy does not cure AIDS
and cannot prevent the spread of HIV, and so it does not hold the promise of a true
vaccine for containing the AIDS epidemic.
Classical vaccines against viruses are often made from the surface proteins of a
virus. These vaccines stimulate the human immune system to recognize viral envelope
proteins as foreign, destroy them, and keep a record of it, so that the immune system
can identify and eradicate the virus in a later encounter.
However, HIV viral envelope proteins are extremely variable because the virus must
mutate rapidly in order to survive (see DETOUR: The Red Queen Effect). The HIV PAGE 229
population in a single infected individual rapidly evolves to evade the human immune
system (Figure 10.1), not to mention that HIV strains taken from different patients
represent multiple highly diverged subtypes. Therefore, a successful HIV vaccine must
be broad enough to account for this variability.
In an effort to counteract HIV’s variability, we could create a single peptide that
contains the least variable segments of the envelope proteins taken from all known
HIV strains and use this peptide as the basis for a universal vaccine fighting all HIV
strains. However, not only do HIV envelope proteins mutate fast, but they are also
“masked” by glycosylation, a post-translational modification that often makes these
proteins invisible to the human immune system (see DETOUR: Glycosylation). As a PAGE 229
result, all attempts at developing an HIV vaccine have thus far failed.
HIV has just nine genes, and in this chapter we will focus on the rapidly mutating
env gene, which has a mutation rate of 1 to 2% per nucleotide per year. The protein
179
CHAPTER 10
VKKLGEQFR-NKTIIFNQPSGGDLEIVMHSFNCGGEFFYCNTTQLFN----------NSTES------DTITL
VKKLGEQFR-NKTIIFNQPSGGDLEIVMHSFNCGGEFFYCNTTQLFN----------NSTDNG-----DTITL
VKKLGEQFR-NKTIIFNQPSGGDLEIVMHSFNCGGEFFYCNTTQLFD----------NSTESNN----DTITL
VDKLREQFGKNKTIIFNQPSGGDLEIVMHTFNCGGEFFYCNTTQLFNSTWNS---TGNGTESYNGQENGTITL
VDKLREQFGKNKTIIFNQPSGGDLEIVMHTFNCGGEFFYCNTTQLFNSTWNG---TNTT--GLDG--NDTITL
VDKLREQFGKNKTIIFNQSSGGDLEIVTHTFNCGGEFFYCNTTQLFNSNWTG---NSTE--GLHG--DDTITL
VKKLGEQFG-NKTIIFNQSSGGGLEIVMHSFNCGGEFFYCNTTQLFNN--TR-----NSTESNNGQGNDTTTL
VKKLREQFGKNKTIIFKQSSGGDLEIVTHTFNCAGEFFYCNTTQLFNSNWTE-----NSITGLDG--NDTITL
VGKLREQFGK-KTIIFNQPSGGDLEIVMHSFNCQGEFFYCNTTRLFNSTWDNSTWNSTGKDKENGN-NDTITL
FIGURE 10.1 A multiple alignment of a short region of gp120 proteins sampled from a
single HIV-positive patient at nine different time points. Almost half of the columns
(shown in darker text) are not conserved across all time points, illustrating how quickly
HIV evolves, even within an individual host. Amino acids differing from the most
common symbol in a column are shown in blue.
encoding the env gene then gets cut into glycoprotein gp120 (approx. 480 amino acids)
and glycoprotein gp41 (approx. 345 amino acids). Together, gp120 and gp41 form the
envelope spike, which mediates entry of the HIV virus into human cells.
Since HIV mutates so fast, different HIV isolates may have different phenotypes,
thus requiring different drug cocktails. For example, HIV viruses can be divided into
fast-replicating syncytium-inducing (SI) isolates and slow-replicating non-syncytium-
inducing (NSI) isolates. During infection, viral proteins like gp120 that are used by
HIV to enter the cell are transported to the cell surface, where they can cause the host
cell membrane to fuse with neighboring cells. This causes dozens of human cells to fuse
their cell membranes into a giant, nonfunctional syncytium, or abnormal multinucleate
cell (Figure 10.2). This mechanism allows an SI virus to kill many human cells by
infecting only one.
FIGURE 10.2 A syncytium with multiple nuclei in an HIV patient.
Because gp120 is important in classifying a virus as SI or NSI, biologists are inter-

ested in determining which amino acids in gp120 can be used for this classification.
180
In 1992, Jean-Jacques De Jong analyzed a multiple alignment of the V3 loop region in

gp120 (Figure 10.3 (top)) and devised the 11/25 rule, which asserts that an HIV strain is
more likely to have an SI phenotype if the amino acid at either positions 11 or 25 of its
V3 loop is arginine or lysine. It later was shown that many other positions influence the
SI/NSI phenotype.
CMRPGNNTRKSIHMGPGKAFYATGDIIGDIRQAHC
CMRPGNNTRKSIHMGPGRAFYATGDIIGDTRQAHC
CMRPGNNTRKSIHIGPGRAFYATGDIIGDIRQAHC
CMRPGNNTRKSIHIGPGRAFYTTGDIIGDIRQAHC
CTRPNNNTRKGISIGPGRAFIAARKIIGDIRQAHC
CTRPNNYTRKGISIGPGRAFIAARKIIGDIRQAHC
CTRPNNNTRKGIRMGPGRAFIAARKIIGDIRQAHC
CVRPNNYTRKRIGIGPGRTVFATKQIIGNIRQAHC
CTRPSNNTRKSIPVGPGKALYATGAIIGNIRQAHC
CTRPNNHTRKSINIGPGRAFYATGEIIGDIRQAHC
CTRPNNNTRKSINIGPGRAFYATGEIIGDIRQAHC
CTRPNNNTRKSIHIGPGRAFYTTGEIIGDIRQAHC
CTRPNNNTRKSINIGPGRAFYTTGEIIGNIRQAHC
CIRPNNNTRGSIHIGPGRAFYATGDIIGEIRKAHC
CIRPNN-TRRSIHIGPGRAFYATGDIIGEIRKAHC
CTRPGSTTRRHIHIGPGRAFYATGNILGSIRKAHC
CTRPGSTTRRHIHIGPGRAFYATGNI-GSIRKAHC
CTGPGSTTRRHIHIGPGRAFYATGNIHG-IRKGHC
CMRPGNNTRRRIHIGPGRAFYATGNI-GNIRKAHC
CMRPGTTTRRRIHIGPGRAFYATGNI-GNIRKAHC
FIGURE 10.3 A multiple alignment of the V3 loop region taken from twenty HIV
patients, and the motif logo of this alignment. The alignment’s 11th and 25th columns
are shown in darker text; occurrences of arginine (R) or lysine (K) in these columns
are shown in red. The 11/25 rule will classify six of the patients as infected with an SI
isolate. Although the V3 loop is an important and rather conserved segment of gp120,
the level of conservation varies across different positions. For example, whereas the first
and last positions are extremely conserved, positions 11 and 25 exhibit high variation.
Limitations of sequence alignment
Before biologists could even start to examine the question of predicting HIV phenotypes
from gp120 sequences, they faced the problem of constructing accurate multiple align-
ments of these sequences. Indeed, a single misalignment, placing an incorrect amino
181
CHAPTER 10
acid at a position influencing the SI/NSI phenotype, could lead to a faulty classification
of HIV phenotypes. And we already know from Chapter 5 that constructing a multiple
alignment of highly diverged sequences is a difficult algorithmic problem.
Figure 10.3 (bottom) shows a motif logo from the V3 loop of gp120 and illustrates
that some positions in gp120 are relatively conserved, whereas others are extremely
variable. Furthermore, this motif logo does not account for insertions and deletions,
which are prevalent in other regions of gp120 that are less conserved than the V3 loop.
These insertions and deletions make analyzing gp120 even more complex.
Because the columns of the multiple alignment in Figure 10.3 have varying levels of
conservation, we question the wisdom of using the same amino acid scoring matrix (as
well as indel penalties) across different columns of an alignment. A better approach
would use a different scoring approach at different columns. For example, an amino
acid differing from R in position 3 of the alignment in Figure 10.3 should incur a larger
penalty than an amino acid deferring from S in position 11.
In other words, the problem formulation of multiple sequence alignment introduced
in Chapter 5 does not offer an adequate translation of the biological problem of HIV
classification into an algorithmic problem. We must therefore devise a new problem
formulation for sequence alignment that will lead to a statistically solid analysis of
gp120 proteins. But first, we will ask you to join us in our time machine for one more trip.
Gambling with Yakuza
The Japanese crime syndicates called yakuza descend from groups of 18th Century
traveling gamblers called bakuto. (In fact, “yakuza” is the name of a losing hand in a
Japanese card game.) One of the most popular games that the bakuto would host in
their makeshift casinos is called Chō-Han. In this game, which literally translates as
“even-odd”, the dealer rolls two dice, and players wager on whether the sum of the dice
will be even or odd.
Although playing Chō-Han in a yakuza gambling house would undoubtedly make
for a fun evening, we can play an equivalent — albeit less exciting — game called
“Heads or Tails” by flipping a coin and wagering on the outcome. Assume that for some
strange reason, more people wager on tails than on heads in this game. Then a crooked
dealer might use a biased coin that is more likely to result in heads than tails. We will
assume that this biased coin results in heads with probability 3/4.
182
STOP and Think: Say that you play Heads or Tails 100 times, and the coin
produces heads 63 times. Is the dealer cheating? Was the coin fair or biased?
This question is not well-formulated, since either coin could have produced any se-
quence of flips. But can we determine which coin is more likely to have been used?
We write the probabilities of tails (“T”) and heads (“H”) for the fair coin (F) as
PrF (“H”) = 1/2 PrF (“T”) = 1/2
and the probabilities for the biased coin (B) as
PrB (“H”) = 3/4 PrB (“T”) = 1/4
Since coin flips are independent events, the probability that n flips of the fair coin will
generate a given sequence x = x1 x2 . . . xn with k occurrences of “H” is
n
Pr( x |F) = ’ PrF (xi ) = (1/2)n .
i =1
On the other hand, the probability that the biased coin will generate the same
sequence is
n
Pr( x | B) = ’ PrB (xi ) = (1/4)n k
· (3/4)k = 3k /4n .
i =1
If Pr( x | F ) > Pr( x | B), then the dealer more likely used a fair coin, and if Pr( x | F ) <
Pr( x | B), then the dealer more likely used a biased coin. The numbers (1/2)n and 3k /4n
are so small for large n that in order to compare them, we will use their log-odds ratio,
✓ ◆ ✓ n◆
Pr( x | F ) 2
log2 = log2 = n k · log2 3.
Pr( x | B) 3k
EXERCISE BREAK: Show that Pr( x | F ) is larger than Pr( x | B) when the log-odds
ratio is positive (i.e., when k/n < 1/ log2 3) and smaller than Pr( x | B) when the
log-odds ratio is negative (i.e., when k/n > 1/ log2 3).
Returning to our example of witnessing k = 63 heads in n = 100 flips, the log-odds

ratio is positive, since
k/n = 0.63 < 1/ log2 3 ⇡ 0.6309.

It follows that Pr( x | F ) > Pr( x | B), and so the dealer most likely used a fair coin, even
though 63 is closer to 75 than it is to 50.
183
CHAPTER 10
Two Coins up the Dealer’s Sleeve
In bakuto gambling houses, a Chō-Han dealer would remove his shirt during play,
displaying a tattooed chest, in order to reduce any suspicions of dice tampering. (These
tattoos would later become a yakuza tradition.) We will assume, however, that in Heads
or Tails, the crooked dealer is wearing a shirt and keeps both coins up a sleeve, secretly
changing them back and forth whenever he likes during the sequence of flips. Since he
does not want to be caught switching coins, he does so only occasionally.
We will assume that the crooked dealer switches coins with probability 0.1 after
each flip. Given a sequence of coin flips, we must determine when the dealer used the
biased coin and when he used the fair coin.
Casino Problem:
Given a sequence of coin flips, determine when the crooked dealer used a fair coin and
when he used a biased coin.
Input: A sequence x = x1 x2 . . . xn of coin flips made by two possible coins

(F and B).
Output: A sequence p = p1 p2 . . . pn , with each pi being equal to either F
or B and indicating that xi is the result of flipping the fair or biased coin,
respectively.
Unfortunately, this problem is poorly stated, since either coin can generate any out-
come. Instead, we need to determine the most likely sequence of coins used by the dealer.
STOP and Think: Can you reformulate the Casino Problem so that it makes
sense?
A well-defined computational problem for finding the most likely sequence of coins
used by the dealer should somehow grade different sequences p as better answers than
others. One approach to guessing the most likely coin the dealer used for each flip
would be to slide a window (of length t < n) along the sequence of flips x = x1 . . . xn
and then calculate the log-odds ratio under each window. If the log-odds ratio of the
window falls below zero, then the dealer most likely used the biased coin inside this
window; otherwise, the dealer most likely used the fair coin.
STOP and Think: Do you see any problems with this method?
184
There are two issues with the window-sliding approach. First, we have no appar-
ent choice for the length of the window. Second, overlapping windows may clas-
sify the same outcome as caused by both the fair and biased coins. For example, if
x = “HHHHHTTHHHTTTTT”, then the window x1 . . . x10 = “HHHHHTTHHH” has
a negative log-odds ratio, and the window x = x6 . . . x15 = “TTHHHTTTTT” has a posi-
tive log-odds ratio. So which coin did the dealer use on the flips x6 . . . x10 = “TTHHH”?
Finding CG-Islands
In the next section, we will improve our method of grading sequences of coin flips. The
solution will lead us to a computational paradigm that has been successfully applied to
a wide array of bioinformatics problems, including HIV comparison. For now, however,
you may still not believe how coin flipping could possibly relate to sequence comparison.
Thus, we will briefly describe a different biological problem that more clearly relates to
our coin flipping analogy.
In the early 20th Century, Phoebus Levene discovered the four nucleotides making
up DNA. At this time, little was known about DNA (Watson & Crick’s double helix
paper was still half a century away). As a result, Levene doubted that DNA could store
genetic information using just a four-letter alphabet, and he hypothesized that DNA
comprised nearly equal amounts of adenine, cytosine, guanine, and thymine.
A century later, we know that complementary nucleotides on opposing strands of
DNA have equal frequencies because of base pairing — ignoring extremely rare base-
pairing errors. However, it is not true that nucleotide frequencies are approximately
equal on a single strand of DNA. For example, different species have widely varying
GC-content, or the percentage of cytosine and guanine nucleotides in a genome. For
example, the human genome’s GC-content is approximately 42%.
After accounting for the human genome’s skewed GC-content, we might expect that
each of the dinucleotides CC, CG, GC, and GG would occur in the human genome
with frequency 0.21 · 0.21 = 4.41%. However, the frequency of CG in the human
genome is only about 1%! This dinucleotide is so rare because of methylation, the
most common DNA modification, which typically adds a methyl group (CH3 ) to the
cytosine nucleotide within a CG dinucleotide. The resulting methylated cytosine has the
tendency to further deaminate into thymine (see DETOUR: DNA Methylation). As a PAGE 229
result of methylation, CG is the least frequent dinucleotide in many genomes.
Nevertheless, methylation is often suppressed around genes in areas called CG-
islands, where CG appears relatively frequently (Figure 10.4). If you were to sequence a
185
CHAPTER 10
mammalian genome that you knew nothing about, perhaps one of the first things you
might do in order to find genes in this genome is look for CG-islands.
STOP and Think: How would you identify CG-islands in a genome?
A C G T A C G T
A 0.053 0.079 0.127 0.036 A 0.087 0.058 0.084 0.061
C 0.037 0.058 0.058 0.041 C 0.067 0.063 0.017 0.063
G 0.035 0.075 0.081 0.026 G 0.053 0.053 0.063 0.042
T 0.024 0.105 0.115 0.050 T 0.051 0.070 0.084 0.084
FIGURE 10.4 Dinucleotide frequencies for a collection of CG-islands (left) and non-CG-
islands (right) in the human genome computed for a single strand of the X chromosome.
Frequencies of CG are shown in red.
A naive approach to search for CG-islands in a genome would slide a window down
the genome, declaring windows with higher frequencies of CG as potential CG-islands.
The disadvantages of this method are analogous to those of using a sliding window to
determine which coin the crooked dealer most likely used at any given point in time.
We do not know how long the window should be, and overlapping windows may
simultaneously classify the same genomic position as belonging to a CG-island and as
not belonging to a CG-island.
Hidden Markov Models
From coin flipping to a Hidden Markov Model
Our goal is to develop a single concept that models both the crooked dealer and the
search for CG-islands in a genome. To this end, we will think about the crooked dealer
not as a human but as a primitive machine. We do not know how this machine is
constructed, but we do know that it proceeds in a sequence of steps; in each step, it is in
one of two hidden states, F and B, and it emits a symbol, “H” or “T”.
After each step, the machine makes two decisions:
• Which hidden state will I move to next?
• Which symbol will I emit in that state?
186
The machine answers the first question by choosing randomly among the F and B states,
with probability 0.9 of remaining in its current state and probability 0.1 of changing
states. The machine answers the second question by choosing between the symbols “H”
and “T” with probabilities that depend on the state it is in. In our coin flipping example,
the probabilities for state F (0.5 and 0.5) differ from the probabilities for state B (0.75
and 0.25). Our goal is to infer the machine’s most likely sequence of states by analyzing
the sequence of symbols that it emits.
We have just transformed the dealer into an abstract machine called a Hidden
Markov Model (HMM). The only difference between our specialized “coin flipping
machine” and the general concept of an HMM is that the latter can have an arbitrary
number of states and may have arbitrary probability distributions governing which
state to move into and which symbols to emit. In general, an HMM (Â, States, Transition,
Emission) is defined by a set of four objects:
• an alphabet S of emitted symbols;
• a set States of hidden states;
• a |States| ⇥ |States| matrix Transition = (transitionl, k ) of transition probabilities,

where transitionl, k represents the probability of moving from state l to state k;
• a |States| ⇥ |S| matrix Emission = (emissionk (b)) of emission probabilities, where

emissionk (b) represents the probability of emitting symbol b from alphabet S when
the HMM is in state k.
For each state l,
Â transitionl, k = 1
all states k
and
Â emissionl (b) = 1 .
all symbols b from S
EXERCISE BREAK: What are S, States, Transition, and Emission for the HMM
modeling the crooked dealer?
187
CHAPTER 10
The HMM diagram
As illustrated in Figure 10.5, an HMM can be visualized using an HMM diagram, a

graph in which every state is represented by a solid node. Solid directed edges connect
every pair of nodes, as well as every node to itself. Each such edge is labeled with the
transition probability of moving from one state to the other (or remaining in the same
state). In addition, the HMM diagram has dashed nodes representing each possible
symbol from the alphabet S and dashed edges connecting each state to each dashed
node. Each such edge is labeled by the probability that the HMM will emit this symbol
while in the given state.
Emission Transmission
H T F B F B
F 1/2 1/2 F 9/10 1/10
B 3/4 1/4 B 1/10 9/10
FIGURE 10.5 The transition and emission probability matrices for the crooked dealer
HMM described by the HMM diagram shown in the center. This HMM has two states
(gray nodes), F and B. In each state, the HMM can emit one of two symbols (dashed
nodes), heads (“H”) or tails (“T”), with the probabilities shown along dashed edges.
Transition probabilities are shown on solid edges; the crooked dealer HMM transitions
between states F and B with probability 1/10 and remains in the same state with
probability 9/10.
A hidden path p = p1 . . . pn in an HMM is the sequence of states that the HMM

passes through; such a path corresponds to a path of solid edges in the HMM diagram.
Figure 10.6 presents an example in which the crooked dealer HMM produces a sequence
of flips x = “THTHHHTHTTH” with hidden path p = FFFBBBBBFFF, i.e., the fair
coin is used for the first three flips and last three flips, and the biased coin is used for
the five intermediate flips.
Reformulating the Casino Problem
We can now rephrase the improperly formulated Casino Problem as finding the most
likely hidden path p for a string x of symbols emitted by an HMM. To solve this problem,
188
i 1 2 3 4 5 6 7 8 9 10 11
x T H T H H H T H T T H
p F F F B B B B B F F F
Pr(pi ! pi +1 ) 1 9 9 1 9 9 9 9 1 9 9
2 10 10 10 10 10 10 10 10 10 10
Pr(xi | pi ) 1 1 1 3 3 3 1 3 1 1 1
2 2 2 4 4 4 4 4 2 2 2
FIGURE 10.6 A sequence x of emitted symbols along with a hidden path p for the
crooked dealer HMM. Pr(pi ! pi +1 ) denotes the probability transitionpi , pi +1 of tran-
sitioning from state pi to pi +1 . Pr(p0 ! p1 ) is set equal to 1/2 to comply with the
assumption that in the beginning, the dealer is equally likely to use the fair or biased
coin. Pr(xi |pi ) denotes the probability that the dealer produced symbol xi from state
pi and is equal to emissionpi (xi ).
we will first consider the simpler problem of computing the probability Pr( x, p ) that an
HMM follows the hidden path p = p1 . . . pn and emits the string x = x1 . . . xn . Note
that
Â Â Pr( x, p ) = 1.
all strings of emitted symbols x all hidden paths p
STOP and Think: What is Pr( x, p ) for the x and p in Figure 10.6?
Each emitted string x has probability Pr( x ), which is independent of the hidden path
taken by the HMM:
Pr( x ) = Â Pr( x, p ).
all hidden paths p
Each hidden path p has probability Pr(p ), which is independent of the string that the
HMM emits:
Pr(p ) = Â Pr( x, p ).
all strings of emitted symbols x
The event “the HMM follows the hidden path p and emits x” can be thought of as a
combination of two consecutive events:
• The HMM follows the path p. The probability of this event is Pr(p ).
• The HMM emits x, given that the HMM follows the path p. We refer to the
probability of this event as the conditional probability of x given p, denoted
Pr( x |p ).
189
CHAPTER 10
Both of these events must occur for the HMM to follow path p and emit string x, which
implies that
Pr( x, p ) = Pr( x |p ) · Pr(p ).
PAGE 230 To learn more about this formula, see DETOUR: Conditional Probability.
To compute Pr( x, p ), we will first compute Pr(p ). As shown in Figure 10.6, we
write Pr(pi ! pi+1 ) to denote the transition probability of the HMM transitioning
from state pi to pi+1 . For simplicity, we assume that in the beginning, the dealer is
equally likely to use the fair or biased coin, an assumption that is modeled by setting
Pr(p0 ! p1 ) = 1/2 in Figure 10.6, where p0 is a “silent” initial state that does not
emit any symbols. The probability of p is therefore equal to the product of its transition
probabilities (purple elements in Figure 10.6),
n n
Pr(p ) = ’ Pr(pi 1 ! pi ) = ’ transitionpi 1 , pi
.
i =1 i =1
Probability of a Hidden Path Problem:

Compute the probability of an HMM’s hidden path.
10A
Input: A hidden path p in an HMM (S, States, Transition, Emission).
Output: The probability of this path, Pr(p ).
Note that we have already computed Pr( x |p ) for the crooked dealer HMM when the
dealer’s hidden path consisted only of B or F, which we wrote as Pr( x | B) and Pr( x | F ),
respectively. To compute Pr( x |p ) for a general HMM, we will write Pr( xi |pi ) to denote
the emission probability emissionpi ( xi ) that symbol xi was emitted given that the HMM
was in state pi (Figure 10.6). As a result, for a given path p, the HMM emits a string x
with probability equal to the product of emission probabilities along that path,
n
Pr( x |p ) = ’ Pr(xi |pi )
i =1
n
= ’ emissionpi ( xi ) .
i =1
190
Probability of an Outcome Given a Hidden Path Problem:

Compute the probability that an HMM will emit a string given its hidden path.
Input: A string x = x1 . . . xn emitted by an HMM (S, States, Transition, 10B

Emission) and a hidden path p = p1 . . . pn .
Output: The conditional probability Pr( x |p ) that x will be emitted given
that the HMM follows the hidden path p.
Returning to our formula for Pr( x, p ), the probability that an HMM follows path p and
emits string x can be written as a product of emission and transition probabilities,
Pr( x, p ) = Pr( x |p ) · Pr(p )

n
= ’ Pr( xi |pi ) · Pr(pi 1 ! pi )
i =1
n
= ’ emissionpi ( xi ) · transitionpi 1 , pi
.
i =1
EXERCISE BREAK: Compute Pr( x, p ) for the x and p in Figure 10.6. Can you
find a better explanation for x = “THTHHHTHTTH” than p = FFFBBBBBFFF?
STOP and Think: Now that you have learned about HMMs, try designing an
HMM that will model searching for CG-islands in a genome. What barriers do
you encounter?
The Decoding Problem
The Viterbi graph
As we stated in the previous section, in both the crooked dealer and CG-island HMMs,
we are looking for the most likely hidden path p for an HMM that emits a string x. In
other words, we would like to maximize Pr( x, p ) among all possible hidden paths p.
191
CHAPTER 10
Decoding Problem:
Find an optimal hidden path in an HMM given a string of its emitted symbols.
Input: A string x = x1 . . . xn emitted by an HMM (S, States, Transition,

Emission).
Output: A path p that maximizes the probability Pr( x, p ) over all possible
paths through this HMM.
In 1967, Andrew Viterbi used an HMM-inspired analog of a Manhattan-like grid to

solve the Decoding Problem. For an HMM emitting a string of n symbols x = x1 . . . xn ,
the nodes in the HMM’s Viterbi graph are divided into |States| rows and n columns
(Figure 10.7 (middle)). That is, node (k, i ) represents state k and the i-th emitted symbol.
Each node is connected to all nodes in the column to its right; the edge connecting
(l, i 1) to (k, i ) corresponds to transitioning from state l to state k (with probability
transitionl,k ) and then emitting symbol xi (with probability emissionk ( xi )). As a result,
every path connecting a node in the first column of the Viterbi graph to a node in the
final column corresponds to a hidden path p = p1 . . . pn .
We assign a weight of
W EIGHTi (l, k ) = transitionpi 1 ,pi

· emissionpi ( xi )
to the edge connecting (l, i 1) to (k, i ) in the Viterbi graph. Furthermore, we define
the product weight of a path in the Viterbi graph as the product of its edge weights.
For a path from the leftmost column to the rightmost column in the Viterbi graph
corresponding to the hidden path p, this product weight is equal to the product of n 1
terms,
n n 1
’ transitionpi 1 , pi
· emissionpi ( xi ) = ’ W EIGHTi (l, k ).
i =2 i =1
STOP and Think: How does this expression differ from the formula for Pr( x, p )
that we derived in the previous section?
The only difference between the above expression and the expression that we obtained
for Pr( x, p ),
n
’ transitionpi 1 , pi
· emissionpi ( xi ) ,
i =1
192
C B
A A A A A A
|States| B B B B B B
C C C C C C
Number of emitted symbols (n)
A A A A A A
|States| B B B B B B
C C C C C C
FIGURE 10.7 (Top) The diagram of an HMM with three states (emission/transmission
probabilities as well as nodes corresponding to emitted symbols are omitted). (Middle)
Given a string of n symbols x = x1 . . . xn emitted by an HMM, Viterbi’s Manhattan
is a grid with |States| rows and n columns in which each node is connected to every
node in the column to its right. The weight of the edge connecting (l, i 1) to (k, i)
is WEIGHTi (l, k) = transitionl, k · emissionk (xi ). Unlike the alignment graphs from
Chapter 5, in which the set of valid directions was restricted to south, east, and southeast
edges, every node in a column is connected by an edge to every node in the column
to its right in the Viterbi graph. (Bottom) The Viterbi graph with additional source node
(blue) and sink node (red). A path of largest product weight connecting the source to
the sink corresponds to an optimal hidden path solving the Decoding Problem.
is the single factor transitionp0 , p1 · emissionp1 ( x1 ), which corresponds to transitioning

from the initial state p0 to p1 and emitting the first symbol. To model the initial
state, we will add a source node source to the Viterbi graph and then connect source
to each node (k, 1) in the first column with an edge of weight W EIGHT0 (source, k) =
transitionp0 , k · emissionk ( x1 ). We will also assume that the HMM has another silent
193
CHAPTER 10
terminal state that the HMM enters when it has finished emitting symbols. To model
the terminal state, we add a sink node sink to the Viterbi graph and connect every node
in the last column to sink with an edge of weight 1 (Figure 10.7 (bottom)).
Every hidden path p in the HMM now corresponds to a path from source to sink
in the Viterbi graph with product weight Pr( x, p ). Therefore, the Decoding Problem
reduces to finding a path in the Viterbi graph of largest product weight over all paths
connecting source to sink.
EXERCISE BREAK: Find the maximum product weight path in the Viterbi graph
for the crooked dealer HMM when x= “HHTT”.
The Viterbi algorithm
We will apply a dynamic programming algorithm to solve the Decoding Problem. First,
define sk,i as the product weight of an optimal path (i.e., a path with maximum product
weight) from source to the node (k, i ). The Viterbi algorithm relies on the fact that the
first i 1 edges of an optimal path from source to (k, i ) must form an optimal path from
source to (l, i 1) for some (unknown) state l. This observation yields the following
recurrence:
sk, i = max sl, i 1 · (weight of edge between nodes(l, i 1) and (k, i ))

all states l
= max sl, i 1 · W EIGHTi (l, k)
all states l
= max sl, i 1 · transitionpi 1 , pi
· emissionpi ( xi )
all states l
Since source is connected to every node in the first column of the Viterbi graph,
sk, 1 = ssource · (weight of edge between source and (k, 1))

= ssource · W EIGHT0 (source, k)
= ssource · transitionsource, k · emissionk ( x1 )
In order to initialize this recurrence, we set ssource equal to 1. We can now compute the
maximum product weight over all paths from source to sink as
ssink = max sl, n .

all states l
STOP and Think: How can we adapt our algorithm for finding a longest path in
a DAG to find a path with maximum product weight?
194
How fast is the Viterbi algorithm?
We can interpret the Decoding Problem as yet another instance of the Longest Path in
a DAG Problem from Chapter 5 because the path p maximizing the product weight
’in=1 W EIGHTi (pi 1 , pi ) also maximizes the logarithm of this product, which is equal
to Âin=1 log ( W EIGHTi (pi 1 )). Thus, we can substitute the weights of all edges in the
Viterbi graph by their logarithms. Finding a longest path (i.e. a path maximizing the
sum of edge weights) in the resulting graph will correspond to a path of maximum
product weight in the original Viterbi graph. For this reason, the runtime of the Viterbi
10C
algorithm, which you are now ready to implement, is linear in the number of edges
in the Viterbi graph. The following exercise shows that the number of these edges is
O(|States|2 · n), where n is the number of emitted symbols.
EXERCISE BREAK: Show that the number of edges in the Viterbi graph of an
HMM emitting a string of length n is |States|2 · (n 1) + 2 · |States|.
EXERCISE BREAK: Apply your solution for the Decoding Problem to find CG-
islands in the first million nucleotides from the human X chromosome. To help
you design an HMM for this application, you may assume that transitions from
CG-islands to non-CG-islands are rare, occurring with probability 0.001, and that
transitions from non-CG-islands to CG-islands are even more rare, occurring with
probability 0.0001. How many CG-islands do you find?
In practice, many HMMs have forbidden transitions between some states. For such
transitions, we can safely remove the corresponding edges from the HMM diagram
(Figure 10.8 (left)). This operation results in a sparser Viterbi graph (Figure 10.8 (right)),
which reduces the runtime of the Viterbi algorithm, since the runtime of the algorithm
for finding the longest path in a DAG is linear in the number of edges in the DAG.
EXERCISE BREAK: Let Edges denote the set of edges in the diagram of an HMM
that may have some forbidden transitions. Prove that the number of edges in the
Viterbi graph for this HMM is |Edges| · (n 1) + 2 · |States|.
195
CHAPTER 10
A A A A A A
A B B B B B B B
D C C C C C C C
D D D D D D
FIGURE 10.8 (Left) An HMM diagram for an HMM that has four states with some
forbidden transitions, such as from A to D and from C to itself. Edges corresponding to
forbidden transitions between states are not included in the HMM diagram. (Right) The
Viterbi graph for this HMM emitting a string of length 6.
Finding the Most Likely Outcome of an HMM
Dynamic programming allows us to answer questions about HMMs extending beyond

the most likely hidden path. For example, we have already computed the probability
Pr(p ) of a hidden path p. But what about computing Pr( x ), the probability that the
HMM will emit a string x?
EXERCISE BREAK: Which outcome is more likely in the crooked casino:

“HHTT” or “HTHT”? How would you find the most likely sequence of four
coin flips?
Outcome Likelihood Problem:

Find the probability that an HMM emits a given string.
Input: A string x = x1 . . . xn emitted by an HMM (S, States, Transition,

Emission).
Output: The probability Pr( x ) that the HMM emits x.
STOP and Think: To solve the Outcome Likelihood Problem, you can make a
slight change to the Viterbi recurrence sk, i = maxall states l {sl, i 1 · W EIGHTi (l, k)}.
What is the change?
196
We have already observed that Pr( x ) is equal to the sum of Pr( x, p ) over all hidden
paths p. However, the number of paths through the Viterbi graph is exponential in the
length of the emitted string x, and so we will use dynamic programming to develop a
faster approach to compute Pr( x ).
We denote the total product weight of all paths from source to node (k, i ) in the
Viterbi graph as forwardk, i ; note that forwardsink is equal to Pr( x ). To compute forwardk,i ,
we will divide all paths connecting source to (k, i ) into |States| subsets, where each subset
contains those paths that pass through node (l, i 1) (with product weight forwardl,i 1 )
before reaching (k, i ) for some l between 1 and |States|. Therefore, forwardk, i is the sum
of |States| terms,
forwardk, i = Â forwardl, i 1 · (weight of edge connecting (l, i 1) and (k, i ))

all states l
= Â forwardl, i 1 · W EIGHTi (l, k) .
all states l
Note that the only difference between this recurrence and the Viterbi recurrence,
sk, i = max sl, i 1 · W EIGHTi (l, k) ,

all states l
is that the maximization in the Viterbi algorithm has changed into a summation symbol.
We can now solve the Outcome Likelihood Problem by computing forwardsink , which is
equal to
Â forwardk, n . 10D
all states k
Now that we can compute Pr( x ) for an emitted string x, a natural question is to find
the most likely such string. In the crooked dealer example, this corresponds to finding
the most likely sequence of flips over all possible sequences of fair and biased coins that
the dealer could use.
Most Likely Outcome Problem:

Find a most likely string emitted by an HMM.
Input: An HMM (S, States, Transition, Emission) and an integer n.

Output: A most likely string x = x1 . . . xn emitted by this HMM, i.e., a string
maximizing the probability Pr( x ) that the HMM will emit x.
197
CHAPTER 10
EXERCISE BREAK: Solve the Most Likely Outcome Problem (Hint: You may
need to build a 3-dimensional version of Viterbi’s Manhattan).
Profile HMMs for Sequence Alignment
How do HMMs relate to sequence alignment?
You may still be wondering what in the world HMMs have to do with our original
problem of aligning sequences using a column-specific score. As we will see, HMMs
offer an elegant solution to this problem.
Given a family of related proteins, we can check whether a newly sequenced pro-
tein belongs to this family by constructing pairwise alignments between the newly
sequenced protein and each member of the family. If one of the resulting alignments
scores above some stringent threshold, then we can assume that the new protein belongs
to the family. However, this approach may fail to identify distantly related proteins,
such as gp120 proteins taken from different HIV isolates, since these proteins may have
scores falling below the threshold. If a sequence has weak similarities with many family
members, then it most likely belongs to the family.
The problem, then, is to align a new protein to all members of the family at once.
To do so, we must assume that we have already constructed a multiple alignment of a
family of proteins. Fortunately, it will often be obvious that two proteins come from the
same family (e.g., if the proteins are taken from closely related species). Accordingly,
biologists often start by constructing an alignment of undeniably related proteins, which
are typically easy to align even using the simple multiple alignment methods that we
covered in Chapter 5.
Figure 10.9 (first panel) shows a 5 ⇥ 10 alignment Alignment representing a hypo-
thetical family of proteins. Note that the sixth and seventh columns of this alignment
contain many space symbols and likely do not represent meaningful characteristics of
the family. Accordingly, biologists often ignore columns for which the fraction of space
symbols is greater than or equal to a column removal threshold q. Column removal
results in a 5 ⇥ 8 seed alignment (Figure 10.9 (second panel)).
Given a seed alignment Alignment⇤ representing a family of related proteins, our aim
is to build an HMM that realistically models the propensities of symbols in Alignment⇤
represented by the profile matrix P ROFILE (Alignment⇤ ) (Figure 10.9 (third panel)).
Rather than thinking about aligning the existing seed alignment to a given string
Text (representing a new protein), we will instead think about computing the probability
198
1 2 3 4 5 6 7 8
A C D E F A C A D F
A F D A - - - C C F
Alignment A - - E F D - F D C
A C A E F - - A - C
A D D E F A A A D F
A C D E F A D F
A F D A - C C F
Alignment⇤ A - - E F F D C
A C A E F A - C
A D D E F A D F
A 1 0 0 1/5 0 3/5 0 0
C 0 2/4 0 0 0 1/5 1/4 2/5
PROFILE(Alignment⇤ ) D 0 1/4 3/4 0 0 0 3/4 0
E 0 0 0 4/5 0 0 0 0
F 0 1/4 0 0 1 1/5 0 3/5
M1 M2 M3 M4 M5 M6 M7 M8
FIGURE 10.9 A 5 ⇥ 10 multiple alignment Alignment (first panel), its 5 ⇥ 8 seed

alignment Alignment⇤ (second panel), the profile matrix PROFILE (Alignment⇤ ) of the
seed alignment (third panel), and the diagram of a simple HMM that models this profile
(fourth panel). The seed alignment is obtained from the original alignment by ignoring
poorly conserved columns (shaded gray); in this case, we ignore columns for which
the fraction of space symbols is greater than or equal to the threshold ✓ = 0.35. To
better illustrate the relationship between the alignment and its seed alignment, we have
separated the first five columns in the seed alignment from its last three columns and
numbered these columns above the original alignment. The match states MATCH (i)
are abbreviated as Mi . The HMM only has one possible path; it is initially in state
MATCH (1), the transition probability from state MATCH (i) to state MATCH (i+1) is equal
to 1 for all i, and all other transitions are forbidden. Emission probabilities are equal
to frequencies in the profile, e.g., emission probabilities for M2 are 0 for A, 2/4 for C,
1/4 for D, 0 for E, and 1/4 for F.
that the HMM emits Text. If the HMM is designed well, then the more similar Text is to
the strings in Alignment⇤ , the more likely it will be emitted by the HMM.
We will first construct a simple HMM that treats the columns of Alignment⇤ as k
sequentially linked states called match states (Figure 10.9 (fourth panel)), denoted
199
CHAPTER 10
M ATCH (1), . . . , M ATCH (k ). When the HMM enters state M ATCH (i ), it emits sym-
bol xi with probability equal to the frequency of this symbol in the i-th column of
P ROFILE (Alignment⇤ ). The HMM then moves into state M ATCH (i + 1) with transition
probability equal to 1.
The similarity score between Alignment⇤ and Text is the probability Pr(Text) that
the HMM for Alignment⇤ emits Text. This score is equal to the product of frequen-
cies in P ROFILE (Alignment⇤ ) corresponding to each symbol of Text. For example, the
probability that the HMM in Figure 10.9 emits ADDAFFDF is
1
1· 4 · 34 · 15 · 1 · 15 · 34 · 3
5 = 0.003375.
STOP and Think: What are the limitations of the HMM in Figure 10.9?
The HMM that we have proposed does score each column in Figure 10.9 differently,
and to a degree, the more similar Text is to Alignment⇤ , the higher its similarity score.
However, this HMM it is not in keeping with the spirit of HMMs because it has only
one hidden path. Furthermore, it offers a simplistic view of multiple alignment because
it does not account for insertions and deletions. Finally, it can only “align” Text against
Alignment⇤ if the length of Text is exactly equal to the number of columns in Alignment⇤
(Figure 10.10). Yet we will use this limited HMM as the foundation of a more powerful
HMM.
A C D E F A D F
A F D A - C C F
Alignment⇤ A - - E F F D C
A C A E F A - C
A D D E F A D F
Text A D D A F F D F
emission probability 1 1/4 3/4 1/5 1 1/5 3/4 3/5
FIGURE 10.10 Aligning Text = ADDAFFDF against the seed alignment Alignment⇤
represented as a simple HMM in Figure 10.9. This HMM is limited because we are
not able to align a string of length other than 8. Indeed, there is no way to add space
symbols to Text or to add symbols of Text “between” columns of Alignment⇤ .
200
Building a profile HMM
The improved HMM that we propose is called a profile HMM. Given a multiple
alignment Alignment and a column removal threshold q used to obtain a seed alignment
Alignment⇤ , we will denote this profile HMM as HMM (Alignment, q ). Because the
profile HMM will be constructed from the seed alignment, we will also informally
refer to it as HMM (Alignment⇤ ). Given a string Text to align against the existing seed
alignment, our goal is to find an optimal hidden path in the profile HMM by solving
the Decoding Problem for this HMM and the emitted string Text.
As with our first attempt at an HMM from Figure 10.9, the profile HMM will still
traverse its states in an order consistent with traversing the columns of Alignment⇤ from
left to right. However, to align strings Text of varying lengths, we will need more states
in addition to the k match states.
First, we add k + 1 insertion states, denoted I NSERTION (0), . . . , I NSERTION (k) (Fig-
ure 10.11). Entering I NSERTION (i ) allows the profile HMM to emit an additional symbol
after visiting the i-th column of P ROFILE (Alignment⇤ ) and before entering the (i + 1)-
th column. Thus, we will connect M ATCH (i ) to I NSERTION (i ) and I NSERTION (i ) to
M ATCH (i + 1). Furthermore, to allow for multiple inserted symbols between columns
of P ROFILE (Alignment⇤ ), we will connect I NSERTION (i ) to itself.
I0 I1 I2 I3 I4 I5 I6 I7 I8
M1 M2 M3 M4 M5 M6 M7 M8
FIGURE 10.11 An HMM diagram for the seed alignment in Figure 10.9 with match and
insertion states, abbreviated as M and I, respectively. The states I0 and I8 model inser-
tions of symbols occurring before the beginning and end of Alignment⇤ , respectively.
STOP and Think: Can we use the HMM in Figure 10.11 to align a string Text of
length less than 8?
After modeling insertions of new symbols in P ROFILE (Alignment⇤ ), we should also

model “deletions” allowing the profile HMM to skip columns of P ROFILE (Alignment⇤ ).
One way of modeling these deletions is to add edges connecting every state in the
profile HMM to every state on its right (Figure 10.12).
201
CHAPTER 10
I0 I1 I2 I3 I4 I5 I6 I7 I8
M1 M2 M3 M4 M5 M6 M7 M8
FIGURE 10.12 By adding edges connecting each state in the profile HMM from Fig-
ure 10.11 to every state on its right, we can skip columns of Alignment when comparing
Text against this alignment. The above HMM diagram highlights all edges leading into
and out of MATCH (4).
STOP and Think: Revisit the Exercise Break on page 195 to recall that the running
time of the Viterbi algorithm is proportional to the number of edges (with non-
zero transition probabilities) in the HMM diagram. How many edges will the
diagram in Figure 10.12 have? How can we reduce the number of edges in the
HMM diagram?
Instead of skipping states as in Figure 10.12, we can reduce the number of edges in the
HMM diagram by introducing k silent deletion states D ELETION (1), . . . , D ELETION (k)
(Figure 10.13). For example, instead of jumping from M ATCH (i 1) to M ATCH (i + 1),
we can make the transition M ATCH (i 1) ! D ELETION (i ) ! M ATCH (i + 1). Entering
D ELETION (i ) allows the HMM to skip over a column of the alignment without emitting
a symbol.
D1 D2 D3 D4 D5 D6 D7 D8
I0 I1 I2 I3 I4 I5 I6 I7 I8
M1 M2 M3 M4 M5 M6 M7 M8
FIGURE 10.13 Adding silent deletion states (abbreviated as Di ) to the profile HMM
diagram.
STOP and Think: Is the HMM in Figure 10.13 now adequate, or is there anything
else that we have forgotten to add?
202
We can now transition back and forth between match states and insertion states, as
well as back and forth between match states and deletion states, but we cannot tran-
sition between insertion states and deletion states. The profile HMM diagram should
therefore include edges connecting I NSERTION (i ) to D ELETION (i + 1) and connect-
ing D ELETION (i ) to I NSERTION (i ) for each i. As a result, the profile HMM can move
from any match/insertion state to any other match/insertion state on its right by
sidetracking through intermediate deletion states. We obtain the complete profile
HMM diagram shown in Figure 10.14 after connecting the initial state (S) to the first
match/insertion/deletion states and connecting the final match/insertion/deletion
states to the terminal state (E).
D1 D2 D3 D4 D5 D6 D7 D8
S I0 I1 I2 I3 I4 I5 I6 I7 I8 E
M1 M2 M3 M4 M5 M6 M7 M8
FIGURE 10.14 Adding transitions from insertion states to deletion states and vice-versa
completes the profile HMM diagram for the profile matrix from Figure 10.9. Silent
initial and terminal states are shown by S and E, respectively.
STOP and Think: Consider the following questions.
• How many edges does the HMM diagram in Figure 10.14 have? How does
this compare to the HMM diagram in Figure 10.12?
• What does the Viterbi graph of the profile HMM in Figure 10.14 look like?
How many nodes and edges does it have?
Transition and emission probabilities of a profile HMM
In Figure 10.15, we return to the multiple alignment Alignment from Figure 10.9 and
represent each of the five colored rows of this alignment as a path in the diagram of
HMM (Alignment⇤ ). Symbols in the seed alignment Alignment⇤ (non-shaded columns)
correspond to either a match state (non-space symbols) or a deletion state (space sym-
bols). As for symbols not present in the seed alignment (shaded columns), space
symbols are ignored, and non-space symbols are emitted from insertion states.
203
CHAPTER 10
A C D E F AC A D F
A F D A - -- C C F
Alignment A - - E F D- F D C
A C A E F -- A - C
A D D E F AA A D F
D1 D2 D3 D4 D5 D6 D7 D8
S I0 I1 I2 I3 I4 I5 I6 I7 I8 E
M1 M2 M3 M4 M5 M6 M7 M8
A C D E F AC A D F
D1 D2 D3 D4 D5 D6 D7 D8
S I0 I1 I2 I3 I4 I5 I6 I7 I8 E
M1 M2 M3 M4 M5 M6 M7 M8
A F D A (-) C C F
D1 D2 D3 D4 D5 D
D66 D7 D8
S I0 I1 I2 I3 I4 I5 I6 I7 I8 E
M1 M2 M3 M4 M5 M6 M7 M8
A (-) (-) E F D F D C
D1 D2 D3 D4 D5 D6 D7 D8
S I0 I1 I2 I3 I4 I5 I6 I7 I8 E
M1 M2 M3 M4 M5 M6 M7 M8
A C A E F A (-) C
D1 D2 D3 D4 D5 D6 D7 D8
S I0 I1 I2 I3 I4 I5 I6 I7 I8 E
M1 M2 M3 M4 M5 M6 M7 M8
A D D E F AA A D F
FIGURE 10.15 Five paths through the profile HMM corresponding to the five rows in
the alignment from Figure 10.9. Space symbols below an HMM diagram correspond to
deletion states and are shown in parentheses to indicate that they are not emitted by
the HMM.
204
STOP and Think: How would you assign transition and emission probabilities
for the profile HMM of the alignment in Figure 10.15?
To assign the transition probability transitionl, k , we simply take the frequency of transi-
tions from state l to state k made by these colored paths with respect to all paths that
visited state l. For example, in Figure 10.15, four of the colored paths visit M ATCH (5).
Three of these paths then transition to I NSERTION (5), and one transitions to M ATCH (6).
Thus, we set the following transition probabilities leaving M ATCH (5):
transitionM ATCH(5), I NSERTION(5) = 3/4

transitionM ATCH(5), M ATCH(6) = 1/4
transitionM ATCH(5), D ELETION(6) = 0
We can define the transition probabilities from the initial state analogously. For
the multiple alignment in Figure 10.15, we enter M ATCH (1) with probability 1; for a
general profile HMM, the only other states we could enter from the initial state are
I NSERTION (0) and D ELETION (1). The complete matrix of transmission probabilities is
shown in Figure 10.16.
STOP and Think: Due to the small number of strings in the alignment from
Figure 10.15, many of the transition probabilities in the gray cells in Figure 10.16
are equal to zero. What are the possible negative consequences of these zeroes,
and how would you address these consequences?
To assign the emission probability emissionk (b), we divide the number of times that
symbol b was emitted from state k by the total number of symbols emitted from state
k. For example, in Figure 10.15, there are three occurrences of A, one occurrence of
C, and one occurrence of D emitted from the state I NSERTION (5). Also, there are two
occurrences of C, one occurrence of D, and one occurrence of F emitted from M ATCH (2).
We can therefore infer the following emission probabilities for these two states:
emissionI NSERTION(5) (A) = 3/5 emissionM ATCH(2) (A) = 0

emissionI NSERTION(5) (C) = 1/5 emissionM ATCH(2) (C) = 2/4
emissionI NSERTION(5) (D) = 1/5 emissionM ATCH(2) (D) = 1/4
emissionI NSERTION(5) (E) = 0 emissionM ATCH(2) (E) = 0
emissionI NSERTION(5) (F) = 0 emissionM ATCH(2) (F) = 1/4
205
CHAPTER 10
S I0 M1 D1 I1 M2 D2 I2 M3 D3 I3 M4 D4 I4 M5 D5 I5 M6 D6 I6 M7 D7 I7 M8 D8 I8 E
S 1
I0
M1 .8 .2
D1
I1
M2 1
D2 1
I2
M3 1
D3 1
I3
M4 .8 .2
D4
I4
M5 .75 .25
D5 1
I5 .4 .6
M6 .8 .2
D6
I6
M7 1
D7
I7 1
M8 1
D8
I8
E
FIGURE 10.16 The 27 ⇥ 27 matrix of transition probabilities for

HMM(Alignment, 0.35), where Alignment is the multiple alignment from Fig-
ure 10.9. All values in empty cells are equal to zero. Cells shaded gray correspond
to edges in the HMM diagram from Figure 10.14; cells shaded white correspond to
forbidden transitions.
EXERCISE BREAK: Construct the 27 ⇥ 20 emission probability matrix for

HMM (Alignment, 0.35) derived from Alignment in Figure 10.9.
You are now ready to construct the profile HMM for an arbitrary multiple alignment.
Profile HMM Problem:

Construct a profile HMM from a multiple alignment.
10E
Input: A multiple alignment Alignment and a threshold q.
Output: HMM (Alignment, q ).
206
EXERCISE BREAK: Construct a profile HMM for the HIV sequences shown in
Figure 10.1 with q = 0.35.
Classifying proteins with profile HMMs
Aligning a protein against a profile HMM
Given a protein family, represented by Alignment, we can now return to the problem of
deciding whether a newly sequenced protein, represented by Text, belongs to the family.
We first form HMM (Alignment, q ) for some parameter q. As shown in Figure 10.17, a
hidden path through HMM (Alignment, q ) corresponds to a sequence of match, inser-
tion, and deletion states for aligning Text against Alignment.
D1 D2 D3 D4 D5 D6 D7 D8
S I0 I1 I2 I3 I4 I5 I6 I7 I8 E
M1 M2 M3 M4 M5 M6 M7 M8
A C AF D E (-) A (-) F
A C -- D E F AC A D F
A F -- D A - -- C C F
Alignment A - -- - E F D- F D C
A C -- A E F -- A - C
A D -- D E F AA A D F
Text A C AF D E - -- A - F
FIGURE 10.17 (Top) A path through HMM(Alignment, 0.35) for the multiple alignment
from Figure 10.9 and the emitted string Text = ACAFDEAF. (Bottom) The emitted
symbols correspond to aligning Text against Alignment. Specifically, the first two
symbols are emitted from two match states and belong in the first two positions of the
alignment. The next two symbols are emitted from an insertion state and belong in
columns of their own (shown in pink). The space symbols in the seventh and eleventh
columns above correspond to deletion states; these symbols are not emitted by the
HMM. The space symbols in the gray columns do not correspond to any states and are
passed over. The non-shaded columns form an augmented 6 ⇥ 8 seed alignment for
comparison against newly sequenced proteins.
207
CHAPTER 10
To find the “best” alignment of Text against Alignment, we simply need to apply
the Viterbi algorithm to find an optimal hidden path in HMM (Alignment, q ). If the
product weight of this optimal hidden path exceeds a predetermined threshold, then
we may conclude that Text belongs to the protein family, in which case we augment the
existing seed alignment with an additional row corresponding to Text. In this way, we
can recruit more and more distant family members to a seed alignment, adding these
new proteins to the growing multiple alignment, and thus making the resulting profile
HMM more and more suitable for analyzing the protein family of interest.
STOP and Think: If the product weight for a new protein exceeds a threshold
for more than one protein family, how would you classify this protein?
Profile HMMs have finally helped us achieve our original goal of scoring different
columns of a multiple alignment differently based on the frequency of symbols in
each column. For example, say that the seventh column of Alignment⇤ contains more
occurrences of A than C, and the ninth column of Alignment⇤ contains more occurrences
of C than A. A hidden path passing through M ATCH (7) would be rewarded more
for emitting A than C, whereas a hidden path passing through M ATCH (9) would be
rewarded more for emitting C than A.
The return of pseudocounts
The majority of transition probabilities in the gray cells of Figure 10.16 are equal to
zero. (The same is true of emission probabilities.) These zeroes may cause problems;
for example, the path in Figure 10.17 seems perfectly reasonable for Text = ACAFDEAF,
and yet Pr( x, p ) is equal to zero because the transition probability from M ATCH (2) to
I NSERTION (2) for this profile HMM is zero.
As in Chapter 2, we will introduce pseudocounts by adding a small value s to entries
in the transition matrix that correspond to edges of the HMM diagram in Figure 10.14
(i.e., only the gray elements of Figure 10.16). Note that white cells in Figure 10.16,
corresponding to forbidden transitions, are not affected by pseudocounts. The resulting
matrix will then need to be normalized so that the elements in each row sum to 1.
EXERCISE BREAK: Compute the normalized matrix for the matrix in Fig-
ure 10.16 after adding the pseudocount s = 0.01.
We will also add pseudocounts to the matrix of emission probabilities and normalize
the resulting matrix. We refer to the profile HMM defined by the resulting normalized
208
matrices of transition and emission probabilities as HMM (Alignment, q, s).
Profile HMM with Pseudocounts Problem:

Construct a profile HMM with pseudocounts from a multiple alignment.
Input: A multiple alignment Alignment, a threshold value q, and a pseudo- 10F

count value s.
Output: HMM (Alignment, q, s).
STOP and Think: Since the HMM diagram in Figure 10.17 has 25 nodes — not
including the start and end states — the Viterbi graph for the string emitted in
this figure has 25 rows. How many columns does this Viterbi graph have?
We are now ready to align a string Text to a multiple alignment by constructing the
Viterbi graph for this string (Figure 10.18) and solving the Decoding Problem to find
the most likely hidden path.
STOP and Think: Find paths through the Viterbi graph corresponding to the
bottom four hidden paths in Figure 10.15. What happens?
The troublesome silent states
If you reached this point without any questions about Figure 10.18, then we have
successfully concealed from you that solving the Decoding Problem for HMMs with
silent states is not as simple as it may appear: the graph in Figure 10.18 is not a Viterbi
graph! To see why not, consider the path in Figure 10.19, which emits the same string
as Figure 10.18 but passes through one fewer silent deletion state, thus reducing the
number of columns by one. But the Viterbi graph is not allowed to change depending on
the hidden path p, since we know nothing about the hidden path in advance! Instead,
the number of columns in the Viterbi graph must equal the length of the emitted string,
a condition that is violated in both Figure 10.18 and Figure 10.19.
STOP and Think: How can we modify the notion of the Viterbi graph for HMMs
with silent states?
More generally, the Viterbi algorithm does not tolerate silent states other than the initial
and terminal states. In other words, this algorithm assumes that node (k, i ) in the
209
CHAPTER 10
I0 I0 I0 I0 I0 I0 I0 I0 I0
M1 M1 M1 M1 M1 M1 M1 M1 M1
D1 D1 D1 D1 D1 D1 D1 D1 D1
I1 I1 I1 I1 I1 I1 I1 I1 I1
M2 M2 M2 M2 M2 M2 M2 M2 M2
D2 D2 D2 D2 D2 D2 D2 D2 D2
I2 I2 I2 I2 I2 I2 I2 I2 I2
M3 M3 M3 M3 M3 M3 M3 M3 M3
D3 D3 D3 D3 D3 D3 D3 D3 D3
I3 I3 I3 I3 I3 I3 I3 I3 I3
M4 M4 M4 M4 M4 M4 M4 M4 M4
D4 D4 D4 D4 D4 D4 D4 D4 D4
I4 I4 I4 I4 I4 I4 I4 I4 I4
M5 M5 M5 M5 M5 M5 M5 M5 M5
D5 D5 D5 D5 D5 D5 D5 D5 D5
I5 I5 I5 I5 I5 I5 I5 I5 I5
M6 M6 M6 M6 M6 M6 M6 M6 M6
D6 D6 D6 D6 D6 D6 D6 D6 D6
I6 I6 I6 I6 I6 I6 I6 I6 I6
M7 M7 M7 M7 M7 M7 M7 M7 M7
D7 D7 D7 D7 D7 D7 D7 D7 D7
I7 I7 I7 I7 I7 I7 I7 I7 I7
M8 M8 M8 M8 M8 M8 M8 M8 M8
D8 D8 D8 D8 D8 D8 D8 D8 D8
I8 I8 I8 I8 I8 I8 I8 I8 I8
A (-) (-) E F D F D C
FIGURE 10.18 The Viterbi graph for HMM(Alignment, ✓) and a path in this graph
(shown in purple) corresponding to the hidden path for the emitted string AEFDFDC
from Figure 10.15. Edges between columns correspond to allowed transitions in the
HMM diagram from Figure 10.14 and have an implied rightward orientation. Edges
entering nodes corresponding to deletion states are dashed. Emitted symbols are shown
beneath each column.
210
I0 I0 I0 I0 I0 I0 I0 I0
M1 M1 M1 M1 M1 M1 M1 M1
D1 D1 D1 D1 D1 D1 D1 D1
I1 I1 I1 I1 I1 I1 I1 I1
M2 M2 M2 M2 M2 M2 M2 M2
D2 D2 D2 D2 D2 D2 D2 D2
I2 I2 I2 I2 I2 I2 I2 I2
M3 M3 M3 M3 M3 M3 M3 M3
D3 D3 D3 D3 D3 D3 D3 D3
I3 I3 I3 I3 I3 I3 I3 I3
M4 M4 M4 M4 M4 M4 M4 M4
D4 D4 D4 D4 D4 D4 D4 D4
I4 I4 I4 I4 I4 I4 I4 I4
M5 M5 M5 M5 M5 M5 M5 M5
D5 D5 D5 D5 D5 D5 D5 D5
I5 I5 I5 I5 I5 I5 I5 I5
M6 M6 M6 M6 M6 M6 M6 M6
D6 D6 D6 D6 D6 D6 D6 D6
I6 I6 I6 I6 I6 I6 I6 I6
M7 M7 M7 M7 M7 M7 M7 M7
D7 D7 D7 D7 D7 D7 D7 D7
I7 I7 I7 I7 I7 I7 I7 I7
M8 M8 M8 M8 M8 M8 M8 M8
D8 D8 D8 D8 D8 D8 D8 D8
I8 I8 I8 I8 I8 I8 I8 I8
A (-) E F D F D C
FIGURE 10.19 Another path through another “Viterbi graph” emitting the same string
AEFDFDC as in Figure 10.18.
211
CHAPTER 10
Viterbi graph describes the event “the HMM emitted symbol xi when it was in state
k”. However, if k is a silent state, then the role of the node (k, i ) in the Viterbi graph is
poorly defined, as it is unclear how to define the weight of edges entering this node.
Fortunately, we can fix this issue in the case of profile HMMs by defining the Viterbi
graph with |States| rows and |Text| columns (Figure 10.20). Every time the HMM moves
into a deletion state, rather than crossing over to the next column of the Viterbi graph
(as in Figure 10.18 and Figure 10.19), we will move within the same column. When the
HMM moves into a match or insertion state, we will move to the next column. As a
result, every column of the Viterbi graph corresponds to a single emitted symbol, even
though a path can pass through more than one state in a given column.
EXERCISE BREAK: Show that the vertical edge connecting (i, l ) to (i, k), where
k is a deletion state, should be assigned weight equal to transitionl,k .
STOP and Think: Are there any remaining issues with the graph in Figure 10.20?
There is still a minor flaw with the graph in Figure 10.20. If the HMM moves from
the initial state into D ELETION (1), then the HMM will move through the first column
without emitting a symbol. We will therefore transform the initial state into a column of
silent states containing the initial state and all deletion states (Figure 10.21). This way, if
the HMM enters D ELETION (1) from the initial state, it can move downward through
deletion states before transitioning to a match or insertion state in the first column.
You are now ready to use the profile HMM to align a sequence against a seed align-
ment. The only remaining snare is that when computing sk,i — or, equivalently, log (sk,i )
— we must make sure that all incoming scores have been computed. We therefore
suggest the top-down, column-by-column topological ordering for the profile HMM
shown in Figure 10.22.
Sequence Alignment with Profile HMM Problem:

Align a new sequence to a family of sequences using a profile HMM.
10G Input: A multiple alignment Alignment, a threshold q, a pseudocount value
s, and a string Text.
Output: An optimal hidden path emitting Text in HMM (Alignment, q, s).
212
I0 I0 I0 I0 I0 I0 I0
M1 M1 M1 M1 M1 M1 M1
D1 D1 D1 D1 D1 D1 D1
I1 I1 I1 I1 I1 I1 I1
M2 M2 M2 M2 M2 M2 M2
D2 D2 D2 D2 D2 D2 D2
I2 I2 I2 I2 I2 I2 I2
M3 M3 M3 M3 M3 M3 M3
D3 D3 D3 D3 D3 D3 D3
I3 I3 I3 I3 I3 I3 I3
M4 M4 M4 M4 M4 M4 M4
D4 D4 D4 D4 D4 D4 D4
I4 I4 I4 I4 I4 I4 I4
M5 M5 M5 M5 M5 M5 M5
D5 D5 D5 D5 D5 D5 D5
I5 I5 I5 I5 I5 I5 I5
M6 M6 M6 M6 M6 M6 M6
D6 D6 D6 D6 D6 D6 D6
I6 I6 I6 I6 I6 I6 I6
M7 M7 M7 M7 M7 M7 M7
D7 D7 D7 D7 D7 D7 D7
I7 I7 I7 I7 I7 I7 I7
M8 M8 M8 M8 M8 M8 M8
D8 D8 D8 D8 D8 D8 D8
I8 I8 I8 I8 I8 I8 I8
FIGURE 10.20 The Viterbi graph with |States| rows and |Text| columns for the profile
HMM from Figure 10.14 emitting a string Text of length 7 so that edges entering deletion
states are drawn downward within the same column instead of between columns as in
Figure 10.18 and Figure 10.19. The purple path corresponds to the path through the
HMM in Figure 10.15 emitting AEFDFDC.
213
CHAPTER 10
I0 I0 I0 I0 I0 I0 I0
M1 M1 M1 M1 M1 M1 M1
D1 D1 D1 D1 D1 D1 D1 D1
I1 I1 I1 I1 I1 I1 I1
M2 M2 M2 M2 M2 M2 M2
D2 D2 D2 D2 D2 D2 D2 D2
I2 I2 I2 I2 I2 I2 I2
M3 M3 M3 M3 M3 M3 M3
D3 D3 D3 D3 D3 D3 D3 D3
I3 I3 I3 I3 I3 I3 I3
M4 M4 M4 M4 M4 M4 M4
D4 D4 D4 D4 D4 D4 D4 D4
I4 I4 I4 I4 I4 I4 I4
M5 M5 M5 M5 M5 M5 M5
D5 D5 D5 D5 D5 D5 D5 D5
I5 I5 I5 I5 I5 I5 I5
M6 M6 M6 M6 M6 M6 M6
D6 D6 D6 D6 D6 D6 D6 D6
I6 I6 I6 I6 I6 I6 I6
M7 M7 M7 M7 M7 M7 M7
D7 D7 D7 D7 D7 D7 D7 D7
I7 I7 I7 I7 I7 I7 I7
M8 M8 M8 M8 M8 M8 M8
D8 D8 D8 D8 D8 D8 D8 D8
I8 I8 I8 I8 I8 I8 I8
FIGURE 10.21 The final Viterbi graph for the profile HMM in Figure 10.14 emitting
a string of length 7. Edges within the same column have a downward orientation;
edges between columns have a rightward orientation. Once again, the purple path
corresponds to the path through the HMM in Figure 10.15 emitting AEFDFDC.
214
I0 I0 I0 I0 I0 I0 I0
M1 M1 M1 M1 M1 M1 M1
D1 D1 D1 D1 D1 D1 D1 D1
I1 I1 I1 I1 I1 I1 I1
M2 M2 M2 M2 M2 M2 M2
D2 D2 D2 D2 D2 D2 D2 D2
I2 I2 I2 I2 I2 I2 I2
M3 M3 M3 M3 M3 M3 M3
D3 D3 D3 D3 D3 D3 D3 D3
I3 I3 I3 I3 I3 I3 I3
M4 M4 M4 M4 M4 M4 M4
D4 D4 D4 D4 D4 D4 D4 D4
I4 I4 I4 I4 I4 I4 I4
M5 M5 M5 M5 M5 M5 M5
D5 D5 D5 D5 D5 D5 D5 D5
I5 I5 I5 I5 I5 I5 I5
M6 M6 M6 M6 M6 M6 M6
D6 D6 D6 D6 D6 D6 D6 D6
I6 I6 I6 I6 I6 I6 I6
M7 M7 M7 M7 M7 M7 M7
D7 D7 D7 D7 D7 D7 D7 D7
I7 I7 I7 I7 I7 I7 I7
M8 M8 M8 M8 M8 M8 M8
D8 D8 D8 D8 D8 D8 D8 D8
I8 I8 I8 I8 I8 I8 I8
FIGURE 10.22 The topological order of the Viterbi graph from Figure 10.21, which
proceeds top-down and column-by-column.
215
CHAPTER 10
EXERCISE BREAK: Solve the Sequence Alignment with Profile HMM Problem
for the profile HMM you constructed for the multiple alignment in Figure 10.1
along with a gp120 protein taken from chimpanzee Simian Immunodeficiency
Virus (SIV).
STOP and Think: How would you build the Viterbi graph for an arbitrary HMM
with silent states? In which situations will it be impossible to construct a Viterbi
graph of an HMM with silent states?
Are profile HMMs really all that useful?
The Viterbi algorithm applies for any HMM, but we will describe how it works for
profile HMMs in order to make an important point. Define s M ATCH( j), i as the probability
of the most likely hidden path for the prefix x1 . . . xi of x that ends at state M ATCH ( j),
and define s I NSERTION( j), i and s D ELETION( j), i analogously. Because there are only three
edges entering M ATCH ( j), the Viterbi recurrence states that
8
>
<s M ATCH( j 1), i 1 · W EIGHTi ( M ATCH ( j 1), M ATCH ( j))
>
s M ATCH( j), i = max s I NSERTION( j 1), i 1 · W EIGHTi ( I NSERTION ( j 1), I NSERTION ( j))
>
>
:s · W EIGHT ( D ELETION ( j 1), D ELETION ( j))
D ELETION ( j 1), i 1 i
After taking the logarithm of both sides, the resulting recurrence is very similar to the
standard recurrence relation for global pairwise alignment because it is the maximum
of three sums:
⇣ ⌘
log s M ATCH( j), i =
8
>
<log (s M ATCH( j 1), i 1 ) + log ( W EIGHTi ( M ATCH ( j 1), M ATCH ( j)))
>
max log (s I NSERTION( j 1), i 1 ) + log ( W EIGHT i ( I NSERTION ( j 1), I NSERTION ( j)))
>
>
:log (s
D ELETION ( j 1), i 1 ) + log ( W EIGHT i ( D ELETION ( j 1), D ELETION ( j)))
Figure 10.23 shows how a path in a Manhattan-like alignment graph corresponds to

a path through the profile HMM. Diagonal edges, vertical edges, and horizontal edges
in the Manhattan-like graph correspond to match states, insertion states, and deletion
states, respectively.
Figure 10.23 may make it seem that we have wasted your time introducing HMMs,
since it appears that a profile HMM is somehow equivalent to pairwise sequence align-
ment. However, keep in mind that the choice of edges in Figure 10.23 is based on
216
varying transition and emission probabilities. By deriving individual scoring parame-

ters for each column in the alignment matrix, profile HMMs allow us to capture subtle
similarities that can fly under the radar of the simple scoring approaches from Chapter 5.
D1 D2 D3 D4 D5 D6 D7 D8
S I0 I1 I2 I3 I4 I5 I6 I7 I8 E
M1 M2 M3 M4 M5 M6 M7 M8
A C AF D E (-) A (-) F
A
M1
C
M2
I2 A
I2 F
D
M3
E
M4 (-)
D5 A
M6
(-)
D7 F
M8
FIGURE 10.23 (Top) The hidden path through the profile HMM in Figure 10.17 (top)
emitting ACAFDEAF. (Bottom) The path through a Manhattan-like graph corresponding
to this hidden path.
Learning the Parameters of an HMM
Estimating HMM parameters when the hidden path is known
Thus far, our analysis has assumed that we know the parameters of an HMM, i.e., its
transition and emission probabilities. We have described a naive — and not necessarily
217
CHAPTER 10
optimal — heuristic for choosing these parameters for a profile HMM, but it is not clear
how to select parameters for an arbitrary HMM.
Indeed, the largest complication when modeling biological problems with an HMM
is estimating the HMM’s parameters from data. In terms of the crooked casino, imagine
that you know that the dealer is using two coins to cheat, but you don’t know the bias
of the coins or the probability that the dealer switches coins at any given time. Can you
infer these parameters just from the sequence of coin flips?
STOP and Think: Say that you observe the sequence of coin flips “HHTHH-
HTHHTTTH”. What is your best guess for the biases of the two coins and for
the probabilities of switching from one coin to the other? Would your best guess
change if you knew that the hidden path were p = FFFBBFFFFFBBB?
We will collectively refer to the matrices Transition and Emission as Parameters. Our
goal is to find Parameters and p when we are only given the emitted string x. We will
work toward this goal by assuming that we are given x as well as either Parameters or
p, and we must infer the remaining component. If x and Parameters are known, then
we can find the most likely hidden path p using the Viterbi algorithm. However, we
have not yet considered how to estimate Parameters if we know x and the hidden path p.
HMM Parameter Estimation Problem:

Find optimal parameters explaining the emitted string and the hidden path of an HMM.
Input: A string x = x1 . . . xn emitted by an HMM with unknown transition

and emission probabilities following a known hidden path p = p1 . . . pn .
Output: A transition matrix Transition and an emission matrix Emission that
maximize Pr( x, p ) over all possible transition and emission matrices.
If we know both x and p, then we can compute empirical estimates for the transition and
emission probabilities using a method similar to one we used for estimating parameters
for profile HMMs. If Tl,k denotes the number of transitions from state l to state k in the
hidden path p, then we can estimate the probability transitionl,k by computing the ratio
of Tl,k to the total number of transitions leaving state l,
Tl,k
transitionl,k = .
Âall states j Tl,j
218
Likewise, if Ek (b) denotes the number of times symbol b is emitted when the hidden
path p is in state k, then we can estimate the probability emissionk (b) as the ratio of
Ek (b) to the total number of emitted symbols from state k,
Ek (b)
emissionk (b) = .
Âall symbols c in the alphabet Ek (c)
It turns out that the above two formulas for computing Transition and Emission result in
10H
parameters solving the HMM Parameter Estimation Problem.
Viterbi learning
If we know x and Parameters, then we can construct the most likely path p by applying
the Viterbi algorithm to solve the Decoding Problem:
( x, ?, Parameters) ! p
On the other hand, if we know x and p, then reconstructing Parameters amounts to

solving the HMM Parameter Estimation Problem:
( x, p, ?) ! Parameters
STOP and Think: What do the expressions ( x, p, ?) ! Parameters and

( x, ?, Parameters) ! p remind you of?
HMM Parameter Learning Problem:

Estimate the parameters of an HMM explaining an emitted string.
Input: A string x = x1 . . . xn emitted by an HMM with unknown transition

and emission probabilities.
Output: A transition matrix Transition and an emission matrix Emission that
maximize Pr( x, p ) over all possible transition and emission matrices and
over all hidden paths p.
Unfortunately, the HMM Parameter Learning Problem is intractable, and so we will

instead develop a heuristic that is analogous to the Lloyd algorithm for k-means cluster-
ing from Chapter 8. In that algorithm, illustrated in Figure 8.12, we iterated two steps,
“From Centers to Clusters”,
219
CHAPTER 10
(Data, ?, Centers) ! HiddenVector,

and “From Clusters to Centers”,
(Data, HiddenVector, ?) ! Centers.

As for HMM parameter estimation, we begin with an initial random guess for
Parameters. Then, we use the Viterbi algorithm to find the optimal hidden path p:
( x, ?, Parameters) ! p
Once we know p, we will question our original choice of Parameters and apply our
solution to the HMM Parameter Estimation Problem to update Parameters based on x
and p:
( x, p, ?) ! Parameters0
We then iterate over these two steps, hoping that the estimated parameters are getting
closer and closer to the parameters solving the HMM Parameter Learning Problem:
( x, ?, Parameters) ! ( x, p, Parameters) ! ( x, p, ?)
! ( x, p, Parameters ) ! ( x, ?, Parameters0 )
0
! ( x, p 0 , Parameters0 ) ! ( x, p 0 , ?)
! ( x, p 0 , Parameters00 ) ! . . .
This approach to learning the HMM’s parameters is called Viterbi learning.
STOP and Think: Can Pr( x, p ) decrease during Viterbi learning? When would
you decide to stop the Viterbi learning algorithm?
Note that we have not specified how Viterbi learning should terminate. In practice,
there are various stopping rules to control its running time. For example, the algorithm
10I
can be stopped if the number of iterations exceeds a predefined threshold or if Pr( x, p )
changes very little from one iteration to another.
Also, because Viterbi learning is dependent on the initial guess for Parameters, it
may become stuck in a local optimum. Like other heuristics, it is often run many times,
retaining the best choice of Parameters.
EXERCISE BREAK: Apply Viterbi learning to learn parameters for an HMM

modeling CG-islands as well as for the profile HMM for the gp120 HIV alignment
in Figure 10.1.
220
Soft Decisions in Parameter Estimation
The Soft Decoding Problem
In Chapter 8, we introduced a “soft” clustering algorithm, based on the more general

expectation maximization algorithm, that relaxed the Lloyd algorithm’s rigid assign-
ment of points to clusters. Analogously, by generating a single optimal hidden path,
the Viterbi algorithm provides a rigid “yes” or “no” answer to the question of whether
an HMM was in state k at time i. But how certain are we that this was the case?
Returning to the crooked casino analogy once more, say that the i-th coin flip is heads.
If this flip occurs in the middle of ten consecutive heads, then you should be relatively
confident that the biased coin was used. But what if, of the ten flips surrounding the
i-th flip, six are heads and four are tails? In this case, you should be less certain that the
biased coin was used.
In the case of an arbitrary HMM, we would like to compute the conditional proba-
bility Pr(pi = k| x ) that the HMM was in state k at time i given that it emitted string x.
Soft Decoding Problem:

Find the probability that an HMM was in a particular state at a particular moment given
its emitted string.
Input: A string x = x1 . . . xn emitted by an HMM.

Output: The conditional probability Pr(pi = k| x ) that the HMM was in state
k at step i given that it emitted x.
The unconditional probability that a hidden path will pass through state k at time i and
emit x can be written as the sum
Pr(pi = k, x ) = Â Pr( x, p ) .
all paths p with pi =k
The conditional probability Pr(pi = k| x ) is equal to the proportion of paths that pass
through state k at time i and emit x with respect to all paths emitting x:
Pr(pi = k, x )
Pr(pi = k| x ) =
Pr( x )
Âall paths p with pi =k Pr( x, p )
= .
Âall paths p Pr( x, p )
221
CHAPTER 10
STOP and Think: If the Viterbi algorithm for the crooked casino emits a path
p = p1 p2 . . . pn with pi = B, is the dealer more likely to have used a biased
coin at step i? Is it possible that pi = B but that Pr(pi = B| x ) is smaller than
Pr(pi = F | x )?
The forward-backward algorithm
We note that Pr(pi = k, x ) is equal to the sum of product weights Pr(p, x ) of all paths
p through the Viterbi graph for x that pass through the node (k, i ). As shown in
Figure 10.24 (top), we can break each such path into a blue subpath from source to (k, i ),
which we denote pblue , and a (red) subpath from (k, i ) to sink, which we denote pred .
Writing W EIGHT(pblue ) and W EIGHT(pred ) as the respective product weights of these
subpaths yields the recurrence
Pr(pi = k, x ) = Â Pr( x, p )
all paths p with pi =k
= Â Â W EIGHT (pblue ) · W EIGHT (pred )

all paths pblue all paths pred
= Â W EIGHT (pblue ) · Â W EIGHT (pred ) .

all paths pblue all paths pred
We have already computed the sum of product weights of all blue subpaths; it is
just forwardk,i , which we encountered when solving the Outcome Likelihood Problem.
Now we would like to compute the sum of product weights of all red subpaths, which
we denote as backwardk,i , so that the preceding equation becomes
Pr(pi = k, x ) = forwardk,i · backwardk,i .
The name of backwardk,i derives from the fact that to compute this value, we can
simply reverse the directions of all edges in the Viterbi graph (Figure 10.24 (bottom))
and apply the same dynamic programming algorithm used to compute forwardk,i . Since
the reversed edge connecting (l, i + 1) to (k, i ) has weight W EIGHTi (k, l ) = transitionk,l ·
emissionl ( xi+1 ), we have that
backwardk,i = Â backwardl,i+1 · W EIGHTi (k, l ).

all states l
STOP and Think: How should this recurrence be initialized?
222
A A A A A A
B B B B B B
(k, i)
C C C C C C
A A A A A A
B B B B B B
(k, i)
C C C C C C
FIGURE 10.24 (Top) Each path from source to sink passing through the (black) node
(k, i) in the Viterbi graph from Figure 10.7 (bottom) can be partitioned into two subpaths,
one from source to (k, i) (shown in blue) and another from (k, i) to sink (shown in
red). (Bottom) A “reversed Viterbi graph” in which all edges have been reversed, with
a path from sink to (k, i) highlighted in red. The recurrence for backwardk,i is based
on computing backwardl,i+1 for every state l.
The resulting dynamic programming approach for computing Pr(pi = k, x ) is called

the forward-backward algorithm. Combining the forward-backward algorithm with
our solution to the Outcome Likelihood Problem for computing Pr( x ) yields that
Pr(pi = k, x ) forwardk,i · backwardk,i

Pr(pi = k | x ) =
Pr( x )
=
forward(sink)
, 10J
and so we are ready to solve the Soft Decoding Problem.
EXERCISE BREAK: Consider the following questions.
• For the crooked dealer HMM, compute Pr(pi = k| x ) for x = “THTHH-

HTHTTH” and each value of i. How does your answer change if x =
“HHHHHHHHHHH”?
• Apply your solution for the Soft Decoding Problem to find CG-islands in
the first million nucleotides from the human X chromosome. How does
your answer differ from the solution given by the Viterbi algorithm?
223
CHAPTER 10
We have just seen how to compute the conditional probability Pr(pi = k| x ) that the
HMM passes through node (k, i ) in the Viterbi graph given that the HMM emits x. But
what about the conditional probability Pr(pi = l, pi+1 = k| x ) that the HMM passes
through the edge connecting (l, i ) to (k, i + 1) given that the HMM emits x? As with the
forward-backward algorithm, we can divide every path through the edge in question
into a blue path from source to this edge and a red path from this edge to sink (Fig-
ure 10.25).
EXERCISE BREAK: Prove that Pr(pi = l, pi+1 = k| x ) is equal to forwardl,i ·

W EIGHTi (l, k ) · backwardk,i+1 /forward(sink).
A A A A A A
B B B B B B
(l, i)
C C C C C C
(k, i+1)
FIGURE 10.25 Each path from source to sink in the Viterbi graph passing through the
(black) edge (l, i) ! (k, i + 1) in the Viterbi graph can be partitioned into two subpaths,
one from source to (l, i) (shown in blue) and another from (k, i + 1) to sink (shown in
red).
The probabilities Pr(pi = k | x ) can be put into a |States| ⇥ n responsibility matrix

P⇤ , where P⇤k,i corresponds to a node in the Viterbi graph and is equal to Pr(pi = k| x).
Figure 10.26 (top) shows the “responsibility” matrix P⇤ for the crooked casino.
The probabilities Pr(pi = l, pi+1 = k| x ) can be put into another |States| ⇥ |States| ⇥
(n 1) responsibility matrix P⇤⇤ , where P⇤⇤ l,k,i corresponds to an edge in the Viterbi
graph and is equal to Pr(pi = l, pi+1 = k | x ) (Figure 10.26 (bottom)). For brevity, we
use P to collectively refer to the matrices P⇤ and P⇤⇤ .
EXERCISE BREAK: What is the complexity of an algorithm computing the

matrices P⇤ and P⇤⇤ ?
224
T H T H H H T H T T H
F 0.636 0.593 0.600 0.533 0.515 0.544 0.627 0.633 0.692 0.686 0.609
B 0.364 0.407 0.400 0.467 0.485 0.456 0.373 0.367 0.308 0.314 0.391
1 2 3 4 5 6 7 8 9 10
FF 0.562 0.548 0.507 0.473 0.478 0.523 0.582 0.608 0.643 0.588
FB 0.074 0.045 0.093 0.059 0.037 0.022 0.045 0.025 0.049 0.098
BF 0.031 0.053 0.025 0.042 0.066 0.104 0.051 0.084 0.043 0.022
BB 0.333 0.354 0.374 0.426 0.418 0.351 0.322 0.282 0.265 0.293
FIGURE 10.26 (Top) The responsibility matrix ⇧ ⇤ , where x = “THTHHHTHTTH”
and the emission/transition matrices Parameters are taken from the crooked dealer
⇤
HMM in Figure 10.5. ⇧k,i is equal to Pr(pi = k|x). (Bottom) The responsibility
⇤⇤ ⇤⇤
matrix ⇧ , where ⇧l,k,i = Pr(pi = l, pi+1 = k|x) for the same emitted string and
emission/transition matrices.
Baum-Welch Learning
The expectation maximization algorithm for parameter estimation, called Baum-Welch

learning, alternates between two steps. In the E-step, it estimates the responsibility
profile P given the current parameters:
( x, ?, Parameters) ! P
Then, in the M-step, it re-estimates the parameters from the responsibility profile:
( x, P, ?) ! Parameters
We have already implemented the E-step of the expectation maximization algorithm,

but the question remains how to design the M-step.
When we know the hidden path, the previously defined estimators for Parameters,
reproduced below, define optimal choices for a given hidden path p:
Tl,k Ek (b)
transitionl,k = emissionk (b) = .
Âall states j Tl,j Âall symbols c in the alphabet Ek (c)
Here, Tl,k is the number of transitions from state l to state k in the hidden path p, and
Ek (b) is the number of times symbol b is emitted when the hidden path p is in state k.
EXERCISE BREAK: How would you redefine these estimators when the hidden
path is unknown?
225
CHAPTER 10
To see how to estimate transitionl,k and emissionk (b) when the hidden path is unknown,
we will compute Tl,k and Ek (b) for a known path p in a slightly different way to make
the transition from hard to soft choices more apparent. First, define the following binary
variables:
( (
i =
1 if p i = l and p i + 1 = k 1 if pi = k and xi = b
Tl,k Eki (b) =
0 otherwise 0 otherwise
With this notation, the formulas computing Tl,k and Ek (b) can be rewritten as
n 1 n
Tl,k = Â Tl,ki Ek (b) = Â Eki (b)
i =1 i =1
When the hidden path is unknown, we will substitute the binary variables Tl,k and
i and Ei ( b ) that are computed in terms of the conditional
Ek (b) for new variables Tl,k k
probabilities that a hidden path will pass through a given node or edge of the Viterbi
graph:
i
Tl,k = Pr(pi = l, pi+1 = k| x ) Eki (b) = Pr(pi = k| x )
= P⇤⇤
l,k,i = P⇤k,i if xi = b and 0 otherwise
Armed with these probabilities computed in the previous section, we can compute
new estimates for Parameters that often perform better in practice than the estimates
provided by Viterbi learning:
n 1 n
10K transitionl,k = Â P⇤⇤
l,k,i emissionk (b) = Â P⇤k,i
i =1 i =1
STOP and Think: Should we normalize the transition and emission probabilities
in the above equations? For example, do the above equations imply that all
transition probabilities leaving state l must sum to 1?
EXERCISE BREAK: Use Baum-Welch learning to learn parameters for the HMM
modeling CG-islands and for the HIV profile HMM. Compare these parameters
with parameters derived by applying Viterbi learning.
226
The Many Faces of HMMs
Profile HMMs for multiple sequence alignment and HMMs finding CG-islands are
just two examples of many applications of HMM in bioinformatics. Furthermore,
applications of HMMs to HIV analysis are not limited to profile HMMs but also include
analysis of HIV resistance against antiviral drug therapies.
Early in the chapter, we mentioned that patients infected with HIV are treated with
a cocktail of several drugs. These drugs attempt to suppress replication of the virus,
but HIV often mutates into drug-resistant strains that eventually dominate the virus
population in a host, making a drug cocktail progressively ineffective. HIV viruses are
often sequenced after drug therapy has failed in order to decide how to reformulate the
drug cocktail. Thus, understanding HIV’s pathways to drug resistance is important to
design an effective cocktail.
Yet modeling HIV resistance pathways is a difficult task. Two mutations that are
advantageous for the virus may interact synergistically, causing the double mutation
to be fixed more often than we might predict from the frequencies of the individual
substitutions. Mutations can also interact antagonistically, resulting in mutants that are
less fit than we might predict.
In 2007, Niko Beerenwinkel and Mathias Drton introduced an HMM-based model
for HIV evolution and developing drug resistance. However, their HMM is far too
complex to explain here. We nevertheless mention it here in order to emphasize the
power of HMMs. Even though they may seem like simple machines that flip coins
and emit symbols, HMMs can be applied to tackle complex bioinformatics problems
ranging from gene prediction to regulatory motif finding.
Epilogue: Nature is a Tinkerer and not an Inventor
The sequence of amino acids in a protein encodes its 3-D structure, which often defines
the biological function of a protein. For example, a zinc finger is an element of the 3-D
structure of zinc finger proteins (Figure 10.27). By arranging two cysteines and two
histidines close to each other in a zinc finger protein’s amino acid sequence, the protein
is able to “grab” a zinc ion and fold tightly around it. Zinc fingers are so useful that
they are found in thousands of human proteins. Furthermore, zinc finger proteins are
used for more than binding zinc, as many of these proteins bind to other metals or even
non-metals.
Over 100,000 experimentally determined protein structures are currently known,
but many of them represent very similar structures or share segments with very similar
227
CHAPTER 10
FIGURE 10.27 A zinc ion (shown in green) is held in place by two histidine residues
and two cysteine residues within a zinc finger.
structure. A protein domain is a conserved part of a protein that can function indepen-
dently of the rest of the protein. Domains vary in length, but the average domain length
is approximately 100 amino acids (zinc finger domains are only 20-30 amino acids long.)
Many proteins consist of several domains, and the same domain may appear (with
variations) in many proteins.
Nobel Laureate François Jacob famously said in 1977: “Nature is a tinkerer and not
an inventor.” In accordance with this principle, nature uses domains as building blocks,
shuffling them into different arrangements to create multi-domain proteins. Most
domains once existed as independent proteins; for example, many domains belonging
to human multi-domain proteins can be found as single-domain proteins in bacteria.
Multi-domain proteins occur naturally when a genome rearrangement creates a new
protein-coding sequence containing parts of the coding sequences from two different
genes. Association of two domains into a single protein often provides an evolutionary
advantage, such as when both domains are enzymes, in which case it may be beneficial
for the cell to ensure a fixed one-to-one ratio of the enzymes’ activities.
Since proteins are often built from multiple domains with different structures and
functions, biologists commonly analyze individual domains instead of entire proteins in
order to understand evolutionary relationships. Since sequence similarities between do-
mains with similar structures can be extremely low, classifying domains into structural
families can be difficult. The Pfam database, which contains over 10,000 HMM-derived
multiple alignments of protein domain families, can be used to analyze new protein
sequences.
CHALLENGE PROBLEM: Using the Pfam HMM for gp120 (constructed from
a seed alignment of just 24 gp120 proteins), construct alignments of all known
gp120 proteins and identify the “most diverged” gp120 sequence.
228
Detours
The Red Queen Effect
The Red Queen Effect is the hypothesis that evolution is necessary not only to equip
organisms with an advantage in a fixed environment, but also to help them survive in
response to changing environments. Its name derives from a statement that the Red
Queen made to Alice in Lewis Carroll’s Through the Looking-Glass:
Now, here, you see, it takes all the running you can do, to keep in the same place.
The Red Queen Effect is often seen in predator-prey relationships. For example,
an adaptation may help wolves run a little faster, and caribou must evolve in turn to
survive. The result is that wolves and caribou appear to run at the same speed, with the
slowest wolves are starving and the slowest caribou being eaten.
Glycosylation
Cells have a dense coating of sugar chains, called glycans, on their surface. Glycans are
often post-translational modifications of glycoproteins, which modulate interactions
with other cells in a multicellular organism or between the cell and another organism
(e.g., between human cells and a virus). For example, influenza infection begins with
an interaction between the proteins on the virus’s surface and glycans on the host cell’s
surface.
Glycans are constructed from a family of building blocks called monosaccharides.
Each monosaccharide can be linked with other monosaccharides to form complex,
tree-like structures (Figure 10.28).
DNA methylation
DNA methylation results in the addition of a methyl group (CH3 ) to a cytosine or

guanine nucleotide (Figure 10.29), which often alters the expression of nearby genes.
Genes that acquire a high concentration of methylated residues in their upstream regions
have suppressed expression. DNA methylation is vital to development, and both DNA
hypermethylation and hypomethylation have been linked to various cancers.
DNA methylation is important in the process of cell differentiation, in which embry-
onic stem cells become specialized tissues. The change is often permanent, preventing a
cell from reverting to a stem cell or converting to a different cell type. Methylation is
inherited during cell division but is usually removed during zygote formation.
229
CHAPTER 10
Mannose
GlcNAc
Galactose
Sialic acid
Fucose
CH 2OH
OH O OH
OH
OH
FIGURE 10.28 (Top) Five types of monosaccharides along with three examples of how
these monosaccharides are assembled into glycans in humans. (Bottom) The chemical
formula for the monosaccharide galactose.
NH 2
CHC3
N
N O
H
FIGURE 10.29 DNA methylation of a cytosine nucleotide base, with the methyl group
shown in blue.
Conditional probability
Let’s return to the game of Chō-Han and analyze the sum s of two standard six-sided
dice. Let A be the event that s is odd and B be the event that s is larger than 10. The
probability of A is equal to 1/2 because half of the 36 possible outcomes for rolling two
dice produce an odd sum. The probability of B is equal to 3/36 because there are three
outcomes (5 + 6, 6 + 5, and 6 + 6) for which s > 10.
230
STOP and Think: If we tell you that s is larger than 10 (but you cannot see the
dice), is s more likely to be odd or even?
The conditional probability of event A given event B, denoted Pr( A| B), is the proba-
bility that event A will occur given that event B has occurred. For the dice example,
since B corresponds to two tosses with s odd (6 + 5 and 5 + 6) and one toss with s even
(6 + 6), Pr( A| B) = 2/3. Note that Pr( A| B) is completely different from the probability
Pr( A, B) that both events A and B will occur, which is equal to 2/36 (A and B only
occur together for the sums 6 + 5 and 5 + 6).
More generally, the conditional probability Pr( A| B) is often defined using the fol-
lowing formula:
Pr( A, B)
Pr( A| B) = .
Pr( B)
EXERCISE BREAK: To test your knowledge of conditional probability, consider

the following puzzle, called the “Monty Hall Problem”, which originally ap-
peared in a letter to American Statistician in 1975 and has stumped many aspiring
mathematicians over the years:
Suppose you’re on a game show, and you’re given the choice of three doors:
Behind one door is a car; behind the others, goats. You pick a door, say No. 1,
and the host, who knows what’s behind the doors, opens another door, say
No. 3, which has a goat. He then says to you, “Do you want to pick door
No. 2?” Is it to your advantage to switch your choice?
231
CHAPTER 10
Bibliography Notes
The decoding algorithm was developed by Viterbi, 1967. The first algorithms for HMM
parameter estimation were developed by Baum et al., 1970. Churchill, 1989, Krogh et al.,
1994, and Baldi et al., 1994 pioneered the application of HMMs in computational biology.
Bateman et al., 2002 described applications of profile HMM alignments for developing
a database of protein domain families Pfam. De Jong et al., 1992 discovered the 11/25
rule. Beerenwinkel and Drton, 2007 constructed an HMM for analyzing HIV resistance.
232
233
WA S T . R E X J U S T A B I G C H I C K E N ?
Paleontology Meets Computing
Growing up in Montana in the 1950s, Jack Horner was shy and introverted. He pro-
gressed so slowly in reading and mathematics that other kids called him stupid. How-
ever, his high school project on dinosaurs took top honors at the local science fair and
was noticed by a University of Montana professor, who helped Jack gain admission to
the university.
Yet Horner’s grades did not improve in college; after failing five consecutive quar-
ters, he dropped out. Years later, Horner would learn that he suffers from dyslexia, a
developmental disorder that is often characterized by difficulty with reading compre-
hension and mathematics despite normal or above-average intelligence.
Fortunately for Horner, he eventually found his calling. After being drafted during
the Vietnam War and later working as a truck driver, he accepted a job as a technician at
Princeton’s Natural History Museum, where he quickly established a reputation among
his peers as a brilliant researcher. He would go on to become the world’s most famous
paleontologist, providing inspiration for one of the main characters of the bestselling
novel Jurassic Park and advising Steven Spielberg for the film adaptation.
Horner was able to succeed despite dyslexia partly because paleontology has not
traditionally required mathematical fluency. However, Horner’s own student would
show that even paleontology is not immune from computing. In 2000, Horner was
exploring his favorite dinosaur graveyard in Montana and discovered a 68 million
year-old Tyrannosaurus rex leg bone fossil. Three years later, he gave a small chunk of
this fossil to his student, Mary Schweitzer, who dissolved it in a demineralizing bath
to study its components but left it in for too long (remember Alexander Fleming?).
When she returned, all that remained was a fibrous substance. Schweitzer then sent this
material to a mass spectrometrist (John Asara) in the hope of detecting T. rex peptides,
or short protein fragments, which had miraculously survived inside the bone.
In 2007, after analyzing thousands of spectra, Asara and Schweitzer published
a paper in Science announcing the discovery of T. rex peptides that closely matched
chicken peptides. Their result provided the first molecular evidence for the controversial
hypothesis that birds evolved from dinosaurs.
The fact that proteins could survive for millions of years was so amazing that it
led to many grandiose claims. Paleontologist Hans Larsson suggested that dinosaurs
would “enter the field of molecular biology and really slingshot paleontology into the
modern world”. The Guardian projected that “scientists may one day be able to emulate
Jurassic Park by cloning a dinosaur.” Horner himself even published a book called How
235
CHAPTER 11
to Build a Dinosaur, detailing his plan to recreate a dinosaur by genetically modifying

the chicken genome.
Yet some scientists remained skeptical. Whereas previous dinosaur studies did not
require much computation, Asara’s T. rex analysis was powered by an algorithm relying
on complicated statistics. In 2008, Science published rebuttals arguing that Asara and
Schweitzer had failed to prove that some of their peptides are not simply statistical
artifacts. But how can we know which side is correct? In this chapter, we will investigate
the T. rex peptides by delving into some algorithms for analyzing spectra.
Which Proteins Are Present in This Sample?
Only four scientists have ever won two Nobel Prizes. One of them is Frederick Sanger,
whose assembly of the first genome in 1977 we mentioned in Chapter 3. Yet Sanger had
already won his first Nobel prize two decades earlier for determining the sequence of
52 amino acids making up insulin, the protein needed to absorb glucose in the blood.
Similarly to how scientists sequence genomes, Sanger broke multiple molecules of
insulin into short peptides, sequenced these peptides, and then assembled them into
the amino acid sequence of insulin (Figure 11.1).
Although protein sequencing was very difficult in the 1950s, DNA sequencing was
impossible. Today, it has become essentially trivial to generate millions of reads for DNA
sequencing, but protein sequencing remains difficult. For this reason, most proteins
are discovered by first sequencing a genome and then predicting all of the genes that
PAGE 273 this genome encodes (see DETOUR: Gene Prediction). By translating the nucleotide
sequence of each protein-coding gene into an amino acid sequence, biologists derive a
putative proteome of a species, i.e., the set of all its proteins.
However, different cells in an organism express different proteins. For example,
brain cells express proteins giving rise to neuropeptides, whereas other cells do not. An
important problem in the study of proteins, or proteomics, is to identify which specific
proteins are present in each biological tissue under different conditions, and how these
proteins interact.
For example, suppose we are studying the chicken ribosome, a complex molecular
machine consisting of many proteins. Knowing the chicken proteome does not tell us
which specific proteins compose the ribosome complex. Instead, we can isolate the
ribosome, break it apart, and identify which proteins it contains. In practice, merely
confirming that a 10 amino acid-long peptide from a known chicken protein is present
in a sample is usually sufficient to confirm this protein’s presence in the sample. The
236
GIVEECCA
GIVEECCASV
GIVEECCASVC
GIVEECCASVCSL
GIVEECCASVCSLY
SLYELEDYC
ELEDY
ELEDYCD
LEDYCD
EDYCD
FVDEHLCG
FVDEHLCGSHL
HLCGSHL
SHLVEA
VEALY
YLVCG
LVCGERGF
LVCGERGFF
GFFYTPK
YTPKA
GIVECCASVCSLYELEDYCDFVDEHLCGSHLVEALYLVCGERGFFFYTPKA
FIGURE 11.1 The peptide assembly that Frederick Sanger used to determine the amino
acid sequence of insulin.
process of confirming that a peptide from a known proteome is present in a sample is

called peptide identification. But how could we form a T. rex proteome?
Although peptide identification dominates modern proteomics studies, the pro-
teomes of many species, including extinct species like T. rex, remain unknown. In
this case, biologists rely on de novo peptide sequencing, or inferring the amino acid
sequence of a peptide without relying on a proteome, and so this is where we will begin.
Decoding an Ideal Spectrum
You may be experiencing déjà vu, since we already discussed cyclic peptide sequencing
in Chapter 4. So we will first remind you of the basics of mass spectrometry, with an
emphasis on its application to linear peptide sequencing.
Given a large number of identical copies of a peptide in a sample — which will
typically contain millions of cells — a mass spectrometer breaks each copy into two
smaller fragments, where different copies of the same peptide may break differently. For
example, one copy of REDCA may break into RE and DCA, and another may break into
237
CHAPTER 11
RED and CA. The fragments RE and RED are called prefixes of REDCA, whereas DCA
and CA are called suffixes of REDCA. Figure 11.2 shows the integer masses of amino
acids.
G A S P V T C I L N D K Q E M H F R Y W
57 71 87 97 99 101 103 113 113 114 115 128 128 129 131 137 147 156 163 186
FIGURE 11.2 The integer mass table of the 20 standard amino acids (reproduced from
Chapter 4).
The algorithmic question that we first pose is analogous to the one that we asked
about cyclic antibiotics but now applied to linear peptides: if we weigh each prefix and
suffix of an unknown peptide, can we reconstruct the peptide? Given an amino acid
string Peptide, its ideal spectrum, denoted I DEAL S PECTRUM(Peptide), is the collection
of integer masses of all its prefixes and suffixes (Figure 11.3 (top)). Note that an ideal
spectrum may have repeated masses; for example, I DEAL S PECTRUM(GPG) = {0, 57, 57,
154, 154, 211}. We say that an amino acid string Peptide explains a collection of integers
Spectrum if I DEAL S PECTRUM(Peptide) = Spectrum.
Fragment "" R RE RED REDC REDCA EDCA DCA CA A

Mass 0 156 285 400 503 574 418 289 174 71
R E D C
A
0 71 156 174 285 289 400 418 503 574
A
C D E R
FIGURE 11.3 (Top) Masses of prefixes and suffixes of REDCA form

IDEALSPECTRUM(REDCA) = {0, 71, 156, 174, 285, 289, 400, 418, 503, 574}.
(Bottom) The DAG GRAPH (IDEALSPECTRUM (REDCA)) in which each mass in the
spectrum is assigned to a node and two nodes are connected with a directed edge if
the difference in their masses is the mass of an amino acid.
238
Decoding an Ideal Spectrum Problem:

Reconstruct a peptide from its ideal spectrum.
Input: A collection of integers Spectrum.

Output: An amino acid string Peptide that explains Spectrum.
We would like to separate masses in a spectrum into those derived from prefix and
suffix peptides, but it is unclear how to do so. Instead, note that if two masses are “one
amino acid mass apart”, then it is likely that they correspond to two prefixes or two
suffixes that differ by a single amino acid. For example, we would not know that masses
400 and 503 correspond to the prefixes RED and REDC (Figure 11.3 (top)). But we could
hypothesize that because the difference between these masses is 103 (the mass of C),
these masses correspond to prefixes or suffixes differing in a single occurrence of C.
This idea motivates a graph-based approach to solving the Decoding an Ideal Spec-
trum Problem. We represent the masses in a spectrum as a sequence Spectrum of integers
s1 , . . . , sm in increasing order, where s1 is zero and sm is the total mass of the (unknown)
peptide. We define a labeled graph G RAPH (Spectrum) by forming a node for each 11A
element of Spectrum, then connecting nodes si and s j by a directed edge labeled by an
amino acid a if s j si is equal to the mass of a (Figure 11.3 (bottom)). As we assumed
when sequencing antibiotics, we do not distinguish between amino acids having the
same integer masses (i.e., the pairs K/Q and I/L).
EXERCISE BREAK: Prove that for any choice of Spectrum, G RAPH (Spectrum) is
a DAG.
Figure 11.3 (bottom) shows that G RAPH (IdealSpectrum(REDCA)) consists of two paths
connecting source = 0 to sink = sm . Concatenating the amino acids along these paths
spells out REDCA and its reverse ACDER, both of which represent solutions of the De-
coding an Ideal Spectrum Problem. The sequencing approach based on spelling a path
from source to sink in G RAPH (Spectrum) is described by the following pseudocode.
DECODINGIDEALSPECTRUM(Spectrum)
construct GRAPH (Spectrum)
find a path Path from source to sink in GRAPH (Spectrum)
return the amino acid string spelled by labels of Path
239
CHAPTER 11
EXERCISE BREAK: Decode the ideal spectrum {0, 57, 114, 128, 215, 229, 316,
330, 387, 444}.
If you attempted this exercise, then there is a good chance that you found a path corre-
sponding to a peptide with an incorrect spectrum (Figure 11.4). It is true that each pep-
tide explaining Spectrum corresponds to a path from source to sink in G RAPH (Spectrum).
However, not every path from source to sink in this graph corresponds to a peptide
explaining Spectrum; consider, for example, the path spelling out GGDTN in Figure 11.4.
For this reason, we must rewrite the above faulty pseudocode as follows.
DECODINGIDEALSPECTRUM(Spectrum)
construct GRAPH (Spectrum)
for each path Path from source to sink in GRAPH (Spectrum)
11B
Peptide the amino acid string spelled by the edge labels of Path
if IDEALSPECTRUM (Peptide) = Spectrum
return Peptide
N A T T T T A N
G G S S G G
0 57 114 128 215 229 316 330 387 444
K/Q D D K/Q
FIGURE 11.4 The DAG GRAPH (Spectrum) for Spectrum = {0, 57, 114, 128, 215, 229,
316, 330, 387, 444}. Only eight of the 32 paths from source to sink in this graph
correspond to peptides explaining Spectrum).
Although D ECODING I DEAL S PECTRUM solves the Decoding an Ideal Spectrum

Problem, exploring all paths in a DAG may be time-consuming, since the number of
PAGE 274 such paths may be exponential in the number of masses in the spectrum (see DETOUR:
Finding All Paths in a Graph).
240
From Ideal to Real Spectra
We already know from our analysis of antibiotics in Chapter 4 that the realities of peptide
sequencing are more harsh than reconstructing a peptide from an ideal spectrum. After
breaking each copy of a peptide into two smaller fragments, a mass spectrometer ionizes
them, resulting in electrically charged fragment ions. It measures each fragment ion’s
mass-to-charge ratio as well as its intensity, or the number of fragment ions detected
at that mass-to-charge ratio (peptides may break frequently at some bonds and hardly
ever at other bonds). As a result, a spectrum is represented as a collection of peaks in a
chart, where a peak’s x-coordinate represents its mass-to-charge ratio, and its height
represents its intensity (Figure 11.5 (top)).
Modern mass spectrometers have limitations on the range of mass-to-charge ratios
that they can detect, making it difficult to analyze entire proteins by mass spectrometry.
As a result, proteins are usually analyzed by first breaking them into shorter peptides
using enzymes called proteases. The most popular protease used in proteomics, and the
one used in the T. rex study, is called trypsin. This protease typically breaks a protein
after the amino acids R and K and results in peptides of average length 14.
Figure 11.5 shows a mass spectrum for one of the T. rex spectra (henceforth referred
to as DinosaurSpectrum), along with its two putative interpretations, ATKIVDCFMTY
and GLVGAPGLRGLPGK. Once we infer the peptide that generated a given spectrum,
we can annotate the spectrum by establishing a correspondence between peaks in the
spectrum and prefixes/suffixes of the peptide. To comply with standard mass spec-
trometry terminology, a peak annotated as the prefix of length i is labeled bi , and a peak
annotated as the suffix of length i is labeled yi .
EXERCISE BREAK: Which of the two interpretations of DinosaurSpectrum in

Figure 11.6 do you think better explains DinosaurSpectrum?
Sequencing a peptide from its real spectrum is even more difficult than it may already
seem. Mass spectra often have “noisy” peaks that contribute to false masses, which may
have higher intensities than peaks corresponding to true prefixes and suffixes. Since
some peptide bonds hardly ever break, the intensities at different mass-to-charge ratios
may differ by orders of magnitude. As a result, a spectrum may not have peaks corre-
sponding to true prefixes and suffixes. For example, in Figure 11.5 (bottom), there are no
peaks annotated as b5 or y9 in DinosaurSpectrum. For this reason, although we ignored
intensities when sequencing antibiotics, we will take them more seriously in this chapter.
241
CHAPTER 11
100-
!" #$%&$#' ( ' )*) +$,
Intensity
0
-
0 200 400 600 800 1000 1200
m/z
FIGURE 11.5 (Top) The spectrum DinosaurSpectrum. (Middle) The same spec-
trum annotated by ATKIVDCFMTY. (Bottom) The same spectrum annotated by
GLVGAPGLRGLPGK. The peak corresponding to the prefix peptide of length i is
annotated as bi , and the peak corresponding to the suffix peptide of length i is anno-
tated as yi . For example, the peak annotated as b10 corresponds to GLVGAPGLRG, and
the peak annotated as y3 corresponds to PGK. Most annotated peaks have charge +1,
but some (such as the one denoted y12++) have charge +2. The charge of the fragment
ion represented by a given peak in the spectrum is not known in advance but can often
be inferred after the peptide that generated the spectrum has been identified. Only
six peaks in DinosaurSpectrum are annotated by GLVGAPGLRGLPGK; peaks b10 , b11 ,
and b13 are annotated by prefix peptides, and peaks y3 , y4 , and y12 are annotated by
suffix peptides.
242
EXERCISE BREAK: REDCA is one possible peptide that explains some but not
all of the masses in the spectrum shown in Figure 11.6, which has false and
missing masses. Can you find another peptide that explains even more masses in
this spectrum?
C F F
K/Q C
C E N
D
A A
0 71 103 113 142 156 170 174 250 285 289 400 403 413 432 503 574
I/L A G D D R
R V A H Y F
FIGURE 11.6 The DAG GRAPH (Spectrum) constructed from Spectrum = {0, 71, 103,
113, 142, 156, 170, 174, 250, 285, 289, 400, 403, 413, 432, 503, 574} with one
missing mass (418) and eight false masses (shown in green) compared to the ideal
spectrum of REDCA.
The issue of false and missing masses is only one of the many complications in mass
spectrometry. When a mass spectrometer breaks a peptide, small parts of the resulting
fragments may be lost, thus lowering their mass. For example, when breaking REDCA
into RE and DCA, RE might lose a water molecule (H2 O) of mass 1 + 1 + 16 = 18,
and DCA might lose an ammonia molecule (NH3 ) of mass 1 + 1 + 1 + 14 = 17. The
respective integer masses of the resulting fragments will be equal to M ASS (RE) 18
and M ASS (DCA) 17.
Because of the many practical complications of mass spectrometry, we will need to
make some simplifying assumptions in order to move toward a computational problem
modeling peptide sequencing. Instead of trying to account for the great variety of
different fragmentation patterns when reconstructing a peptide, we will treat them
as noise. We will also assume that all peaks have charge +1 and that the spectra are
discretized (i.e., that all masses are integers).
243
CHAPTER 11
Peptide Sequencing
Scoring peptides against spectra
Consider an imaginary world in which peptides are built from just two amino acids,
X and Z, having respective masses 4 and 5. For example, given the peptide XZZXX,
prefixes have masses 4, 9, 14, 18, and 22, while suffixes have masses 22, 18, 13, 8, 4.
Now consider the hypothetical spectrum
(0, 0, 0, 3, 8, 7, 2, 1, 100, 0, 1, 4, 3, 500, 2, 1, 3, 9, 1, 2, 2, 0)
arising from this peptide, where the i-th element of this vector corresponds to the
intensity detected at mass i. The prefixes of XZZXX annotate peaks with intensities 3,
100, 500, 9, and 0, whereas the suffixes annotate peaks with intensities 0, 8, 0, 2, and
1. Our goal is to develop an approach for scoring peptides against spectra in the hope
that we will be able to find a peptide that generated the spectrum by simply finding the
peptide with maximum score against this spectrum.
STOP and Think: How would you score a peptide against a spectrum?
One scoring approach is the intensity count, or the sum of intensities of all peaks
annotated by a peptide. For example, we would score peptide XZZXX against the
spectrum above as the sum of intensities of all peaks annotated by XZZXX, i.e., 3 +
100 + 500 + 9 + 0 + 0 + 8 + 0 + 2 + 1. However, the intensity count does not work well
in practice because peak intensities vary widely. As a result, the tallest peaks in the
spectrum (100 and 500 in our toy example) may dominate the score. Since the highest
peaks may represent noise and since lower intensity peaks often represent correct
prefix/suffix peptides, the intensity count is not a good scoring function in practice.
Another scoring approach, called the shared peaks count, simply counts the num-
ber of “tall” peaks annotated by a peptide, i.e., the annotated peaks with intensities
exceeding a predefined threshold. Taking the intensity threshold 5 in our ongoing
example, the shared peak count is 4, since the prefixes of XZZXX annotate the tall peaks
with intensities 100, 500, and 9, and the suffixes annotate the tall peak with intensity
8. The peptide in Figure 11.5 (middle) has shared peak count equal to 10, whereas the
peptide in Figure 11.5 (bottom) has shared peak count equal to 6.
Although the shared peak count works better than the intensity count in practice, it
is still far from ideal; a better approach would account for intensities of peaks without
letting the tallest peaks dominate the score. To achieve this goal, we will convert
244
peptides and spectra into vectors, then define a scoring function that is the dot product
of these vectors.
First, given an amino acid string Peptide = a1 . . . an of length n, we will represent
!
its prefix masses using a binary peptide vector Peptide with M ASS (Peptide) coordinates.
This vector contains a 1 at each of the n prefix coordinates
M ASS ( a1 ), M ASS ( a1 a2 ), . . . , M ASS ( a1 a2 . . . an ) ,
and it contains a 0 in each of the remaining noise coordinates. The toy peptide
XZZXX, whose prefix masses are 4, 9, 14, 18, and 22, corresponds to the peptide vector
(0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1) of length 22 (Figure 11.7).
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
peptide vector 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 1 0 0 0 1
spectral vector 0 0 0 4 -2 -3 -1 -7 6 5 3 2 1 9 3 -8 0 3 1 2 1 0
FIGURE 11.7 The peptide vector of XZZXX and a hypothetical spectral vector generated
by this peptide (assuming that X and Z have masses 4 and 5, respectively). Prefix
coordinates of the peptide vector are shown in boldface. Amplitudes exceeding the
threshold 3 in the spectral vector are shown in color. These bold entries correspond to
three prefix coordinates (blue) and one noise coordinate (red).
Converting a Peptide into a Peptide Vector Problem:

Convert a peptide into a peptide vector.
11C
Input: An amino acid string Peptide.
!
Output: The peptide vector Peptide.
Since a peptide vector uniquely defines the peptide that it originated from, we will use
the terms “peptide vector” and “peptide” interchangeably.
Converting a Peptide Vector into a Peptide Problem:

Convert a peptide vector into a peptide.
Input: A binary vector P. 11D

Output: A peptide whose peptide vector is equal to P (if such a peptide
exists).
245
CHAPTER 11
Where are the suffix peptides?
You may be wondering why peptide vectors only model prefix peptides, since both
prefix and suffix peptides contribute to spectral annotations. Does it not make more
sense to define the peptide vector of XZZXX as (0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1,
0, 0, 0, 1) in order to reflect both its prefix masses (4, 9, 14, 18, and 22) and suffix masses
(4, 8, 13, 18, 22)?
Indeed, any peak of mass s in a spectrum may be interpreted as either a prefix mass
or suffix mass of an unknown peptide Peptide that generated the spectrum. Moreover,
its twin peak, with mass equal to M ASS(Peptide) - s, may be interpreted as either a mass
of a suffix or a mass of a prefix of the same peptide.
To deal with this uncertainty, mass spectrometrists convert a spectrum Spectrum
!
into a spectral vector Spectrum that consolidates information about the intensities of
each peak and its twin into a single value, called an amplitude, at the coordinate
representing the mass of the hypothetical prefix peptide for these twins (Figure 11.8).
Why? Because algorithms interpreting a consolidated spectrum become easier than
PAGE 275 algorithms attempting to account for both twins (see DETOUR: The Anti-Symmetric
Path Problem). To make matters more complicated, amplitudes in the spectral vectors
also account for intensities of ion fragments with various charges and intensities of ion
fragments with water and ammonia molecule loss.
Figure 11.8 shows a spectral vector constructed from DinosaurSpectrum and illus-
trates that amplitudes in a spectral vector may be negative. Negative amplitudes
typically correspond to positions in spectra without peaks or with low intensity peaks.
The correspondence between intensities in a spectrum and amplitudes in its spectral
vector is complex, but in general, the amplitude at mass i reflects the likelihood that the
PAGE 276 (unknown) peptide that generated the spectrum has a prefix with mass i (see DETOUR:
Transforming Spectra into Spectral Vectors).
!
After a peptide Peptide has been transformed into a peptide vector Peptide =
( p1 , . . . , pm ) and a spectrum Spectrum has been transformed into a spectral vector
!
Spectrum = (s1 , . . . , sm ) of the same length, we define S CORE (Peptide, Spectrum) =
! ! ! !
S CORE (Peptide, Spectrum) as the dot product of Peptide and Spectrum,
S CORE (Peptide, Spectrum) = p1 · s1 + . . . + pm · sm .

! !
Note that S CORE (Peptide, Spectrum) is simply an “amplitude count”, the sum of ampli-
! !
tudes in Spectrum that are “annotated” by Peptide. However, this score does not suffer
from the limitations of the intensity count because we have transformed intensities in a
spectrum into amplitudes in its spectral vector. As a result, high-intensity peaks in the
246
spectrum contribute to the score but do not dominate it. The score of the peptide vector
and spectral vector in Figure 11.7 is 4 + 6 + 9 + 3 + 0 = 22; the score of the peptide
vector and spectral vector in Figure 11.8 is 19.
100-
!" #$%&$#' ( ' )*) +$,
Intensity
0
-
0 200 400 600 800 1000 1200
m/z
30
25
20
15
10
Intensity
5
0
-5
-10
-15
200 400 600 800 1000 1200
Position
FIGURE 11.8 (Top) DinosaurSpectrum reproduced from Figure 11.5. (Bottom) The
spectral vector of DinosaurSpectrum. The positions of ones in the peptide vector of
GLVGAPGLRGLPGK (at coordinates 57, 170, 269, 326, 397, 494, 551, 664, 820, 877,
990, 1087, 1144, and 1272) are shown by red lines. The amplitudes of the spectral
vector at these coordinates are equal to -8, +1, -4, - 6, -6, +3, +1, -4, -8, +18, +11,
-10, -7, and 0, respectively. SCORE(GLVGAPGLRGLPGK, DinosaurSpectrum) is the dot
product of the peptide vector and spectral vector, 8 + 1 4 6 6 + 3 + 1 4
8 + 18 + 11 10 7 + 0 = 19. Because most amplitudes are negative, the fact that
SCORE(GLVGAPGLRGLPGK, DinosaurSpectrum) is negative does not necessarily imply
that the spectral interpretation above is incorrect.
STOP and Think: Can you find a peptide vector that scores higher than 22
against the spectral vector in Figure 11.7?
In the remainder of this chapter, we will work with spectral vectors instead of spec-
!
tra. Given a spectral vector Spectrum, our goal is to find a peptide Peptide maximizing
247
CHAPTER 11
! !
S CORE (Peptide, Spectrum). Since the mass of a peptide and the parent mass of the spec-
trum that it generates should be the same, a peptide vector should have the same length
as the spectral vector under consideration. We will therefore define the score between a
peptide vector and a spectral vector of different length as •.
Peptide Sequencing Problem:

Given a spectral vector, find a peptide with maximum score against this spectrum.
!
Input: A spectral vector Spectrum.
!
Output: An amino acid string Peptide that maximizes S CORE(Peptide,
!
Spectrum) among all possible amino acid strings.
Peptide sequencing algorithm

!
Given a spectral vector Spectrum = (s1 , . . . , sm ), we will construct a DAG on m + 1
nodes, labeled with the integers from 0 (source) to m (sink), and then connect node i to
node j by a directed edge if j i is equal to the mass of an amino acid (Figure 11.9). We
will further assign weight si to node i (for 1  i  m) and assign weight zero to node 0.
X X X
0 0 0 0 4 -2 -3 -1 -7 6 5 3 2 1 9 3 -8 0 3 1 2 1 0
Z Z
FIGURE 11.9 The node-weighted DAG for a spectral vector of length m = 22 and the
amino acid alphabet {X, Z} with respective masses 4 and 5. The path from 0 to m
representing peptide XZZXX (corresponding to the peptide vector (0, 0, 0, 1, 0, 0, 0, 0,
1, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1) with score 0 + 4 + 6 + 9 + 3 + 0 is highlighted.
STOP and Think: How does this DAG compare to the DAG G RAPH (Spectrum)
that we constructed to decode an ideal spectrum?
Any path connecting source to sink in this DAG corresponds to an amino acid string
! !
Peptide, and the total weight of nodes on this path is equal to S CORE (Peptide, Spectrum).
248
We have therefore reduced the Peptide Sequencing Problem to the problem of finding a
11E
maximum-weight path from source to sink in a node-weighted DAG.
STOP and Think: In Chapter 5, we developed an algorithm for finding a path of

maximum weight in an edge-weighted DAG. How can we modify this algorithm
to find a path of maximum weight in a node-weighted DAG?
EXERCISE BREAK: Apply your algorithm for the Peptide Sequencing Problem
!
to DinosaurSpectrum.
!
By applying an algorithm solving the Peptide Sequencing Problem to DinosaurSpectrum,
we find the peptide ATKIVDCFMTY with score 96 (Figure 11.5 (middle)). However,
Asara proposed a different peptide, GLVGAPGLRGLPGK with score -19, which we will
call DinosaurPeptide (Figure 11.5 (bottom)). This peptide has much lower score than
ATKIVDCFMTY; in fact, billions of peptides outscore DinosaurPeptide!
STOP and Think: Why do you think that Asara proposed DinosaurPeptide instead
of the higher-scoring ATKIVDCFMTY?
Peptide Identification
The Peptide Identification Problem
If you followed our struggles to sequence antibiotic peptides, then you will agree that
we should be wary of jumping to the conclusion that the highest-scoring peptide for
DinosaurSpectrum must have generated this spectrum.
Despite many attempts, researchers have still not devised a scoring function that
reliably assigns the highest score to the biologically correct peptide, i.e., the peptide
that generated the spectrum. Fortunately, although the correct peptide often does not
achieve the highest score among all peptides, it typically does score highest among all
peptides limited to the species’s proteome. As a result, we can transition from peptide
sequencing to peptide identification by limiting our search to peptides present in the
proteome, which we concatenate into a single amino acid string Proteome.
249
CHAPTER 11
EXERCISE BREAK: How does the number of all peptides of length 10 (which
we must explore in peptide sequencing) compare with the number of peptides
of length 10 in the human proteome? (Note: there are approximately 20,000
protein-coding genes in the human protein, and the average length of a human
protein is approximately 400 amino acids.)
Peptide Identification Problem:

Find a peptide from a proteome with maximum score against a spectrum.
11F !
Input: A spectral vector Spectrum and an amino acid string Proteome.
!
Output: An amino acid string Peptide that maximizes S CORE(Peptide,
!
Spectrum) among all substrings of Proteome.
STOP and Think: In practice, the input to the Peptide Identification Problem is a
set of proteins rather than a single string Proteome. What are the potential pitfalls
of concatenating all proteins as opposed to analyzing each protein separately?
Identifying peptides in the unknown T. rex proteome
You may be wondering why we have returned to peptide identification, since we do

not know the T. rex proteome. It may therefore seem that we cannot apply an algo-
rithm for the Peptide Identification Problem to the spectra obtained from the T. rex fossil.
STOP and Think: How could we form a protein database to search for T. rex
peptides?
Approximately 90% of proteins making up animal bones are collagen. Dinosaur bones
undoubtedly contained collagen, and there is little chance that other proteins could have
survived for millions of years. Since the amino acid sequences of collagens are often
conserved across different species, Asara reasoned that any proteins that had survived
in the T. rex fossil would likely be similar to collagens from present-day species.
As a sanity check, Asara compared the T. rex spectra against the entire UniProt
database, containing proteins from hundreds of species and totaling almost 200 million
amino acids. He also included some mutated versions of collagens from present-day
species in order to model possible differences between these collagens and collagens
250
in T. rex (we will call the resulting protein database UniProt+). It turned out that most
of the high-scoring peptides identified in this database were chicken collagens, which
supported the hypothesis that birds evolved from dinosaurs.
In fact, DinosaurPeptide is only one mutation apart from a chicken collagen peptide.
But how can we test whether this peptide is the correct interpretation of DinosaurSpec-
trum?
Searching for peptide-spectrum matches
Like peptide sequencing algorithms, peptide identification algorithms may return an

erroneous peptide, particularly if the score of the highest-scoring peptide found in the
proteome is much lower than the score of the highest-scoring peptide over all peptides.
For this reason, biologists usually establish a score threshold and only pay attention
to a solution of the Peptide Identification Problem if its score is at least equal to the
threshold.
Given a set of spectral vectors SpectralVectors, an amino acid string Proteome, and
a score threshold threshold, we will solve the Peptide Identification Problem for each
!
vector Spectrum in SpectralVectors and identify a peptide Peptide having maximum
score for this spectral vector over all peptides in Proteome (ties are broken arbitrar-
! !
ily). If S CORE (Peptide, Spectrum) is greater than or equal to threshold, then we conclude
!
that Peptide is present in the sample and call the pair (Peptide, Spectrum) a peptide-
spectrum match (PSM). The resulting collection of PSMs for SpectralVectors is denoted
PSMthreshold (Proteome, SpectralVectors).
PSM Search Problem:

Identify all peptide-spectrum matches scoring above a threshold for a set of spectra and a
proteome.
Input: A set of spectral vectors SpectralVectors, an amino acid string Proteome,

and an integer threshold.
Output: The set PSMthreshold (Proteome, SpectralVectors).
The following pseudocode solves the PSM Search Problem using an algorithm that
you just implemented to solve the Peptide identification Problem, which we call
P EPTIDE I DENTIFICATION.
251
CHAPTER 11
PSMSEARCH(SpectralVectors, Proteome, threshold)

PSMSet an empty set
!
for each vector Spectrum in SpectralVectors
!
11G Peptide PEPTIDEIDENTIFICATION (Spectrum, Proteome)
!
if SCORE (Peptide, Spectrum) threshold
!
add the PSM (Peptide, Spectrum) to PSMSet
return PSMSet
DinosaurPeptide turned out to be the highest scoring peptide for DinosaurSpectrum among
all peptides in the UniProt+ database. But do the billions of peptides not occurring in
this database that outscore DinosaurPeptide imply that the database that Asara formed
is incomplete and that DinosaurSpectrum arose from a different peptide?
The reality is that the highest scoring peptide in a proteome is commonly outscored
by billions of peptides not belonging to the proteome. However, this phenomenon
does not imply that PSMS EARCH has identified the wrong peptide, because the total
number of peptides with the same mass may be measured in the trillions or even
quadrillions. In other words, the billions of peptides that outscore DinosaurPeptide
represent a small fraction of all peptides having the same mass as this peptide. Thus,
we need to complement PSMS EARCH with an evaluation of the statistical significance
of its identified PSMs.
STOP and Think: Suppose that we search 1,000 spectra from a chicken sample
against the chicken proteome, and identify 100 PSMs whose score surpasses a
threshold. How would you estimate the percentage of erroneous PSMs among
these 100 PSMs?
Peptide Identification and the Infinite Monkey Theorem
False discovery rate
To estimate the number of spurious PSMs in PSMthreshold (Proteome, SpectralVectors), we

will construct a decoy proteome DecoyProteome, a randomly generated amino acid
string having the same length as Proteome (with the probability of generating any amino
acid at each position equal to 1/20). We will then solve the PSM Search Problem for
DecoyProteome instead of Proteome for the same score threshold.
We are not interested in the PSMs identified in the randomly generated decoy
proteome, which are nothing more than biologically irrelevant artifacts. The point
252
is that the number of these PSMs will provide a rough estimate for the number of
erroneous PSMs identified in our biologically relevant search against the real proteome.
We will therefore define the false discovery rate (FDR) of a PSM search as the ratio of
the number of decoy PSMs to the number of PSMs identified with respect to the real
proteome,
| PSMthreshold (DecoyProteome, SpectralVectors)|

.
| PSMthreshold (Proteome, SpectralVectors)|
For example, if a search against Proteome results in 100 PSMs, and a search against
DecoyProteome results in just five PSMs, then the FDR would be 5%, and we would
conclude that approximately 95% of identified PSMs are likely valid. On the other hand,
if searching against DecoyProteome returned close to 100 PSMs, then the FDR would
be close to 1, and we would have a hard time making the argument that any peptides
identified in our search against Proteome are biologically relevant.
EXERCISE BREAK: Estimate the FDR when searching all T. rex spectra against
the UniProt+ database with threshold = 80.
STOP and Think: If the FDR turns out to be very high for a given value of
threshold, can we still find reliable PSMs?
Even if the FDR is high, then we should not conclude that our spectral dataset is
worthless or that we are searching against the wrong protein database. We may simply
have selected the wrong score threshold for analyzing our data, since the FDR can vary
widely depending on the choice of this threshold (Figure 11.10).
For the T. rex spectral dataset, we find 27 PSMs in the amino acid string Proteome
derived from UniProt+ and only one PSM in DecoyProteome with score at least equal to
threshold = 100 (FDR = 3.7%). Unfortunately, we cannot automatically conclude that we
have discovered two dozen dinosaur peptides because many of these PSMs correspond
to common laboratory contaminants. Our goal is to figure out whether the remaining
few, including DinosaurPeptide, indeed correspond to dinosaur peptides.
Specifically, FDR helps us analyze the entire set of identified PSMs for all T. rex
spectra, but what about the statistical significance of an individual PSM? In particular,
!
can we determine whether the PSM (DinosaurPeptide, DinosaurSpectrum), which we will
call DinosaurPSM, is statistically significant? To answer this question, we first need to
quantify what we mean by “statistically significant”.
253
CHAPTER 11
90
80
70
60
Number of 50
PSMs 40
30
20
10
0
0 50 100 150 200 250 300 350
Score
FIGURE 11.10 The number of PSMs identified by searching DinosaurSpectrum against

the amino acid string derived from the UniProt+ database (blue) and a decoy proteome
of the same length (red) depending on the score threshold.
The monkey and the typewriter
Imagine that we have locked you in a room with a monkey and a typewriter. The
monkey quickly tires of your company and starts banging away on the typewriter,
creating strings of symbols (we will assume that the monkey particularly likes the
space bar). As you begin to lose your sanity, you test each new string generated by the
monkey to see if some of them are correctly spelled English words. (Figure 11.11). After
all, according to the infinite monkey theorem, the monkey will eventually type Hamlet
PAGE 278 (see DETOUR: The Infinite Monkey Theorem).
FIGURE 11.11 Searching for William Shakespeare.
254
You would be shocked if the monkey immediately typed “To be, or not to be”. But
would you be as surprised if, after cranking out a million strings, the monkey had typed
a dozen English words?
Given a set of strings Dictionary, we define E (Dictionary, n) as the expected number
of occurrences of strings from Dictionary in a randomly generated string of length n,
where the probabilities of generating every letter at each position of the string are
the same. Let EnglishDictionary denote the set of all English words. If it turns out
that, after typing n symbols, the monkey types significantly more English words than
E (EnglishDictionary, n), then we have every reason to believe that the monkey can
spell! On the other hand, if the monkey types about the same number of words as
E (EnglishDictionary, n), then the monkey is probably not Shakespeare reincarnate.
The Monkey and the Typewriter Problem:

Find the expected number of strings from a dictionary appearing in a randomly generated
text.
Input: A set of strings Dictionary and an integer n.

Output: E (Dictionary, n).
STOP and Think: What do the monkey and the typewriter have to do with mass
spectrometry?
EXERCISE BREAK: What is the expected number of times that the string
SHAKESPEARE appears in a randomly generated English string (with no spaces)
of length 200 million?
Statistical significance of a peptide-spectrum match
Now imagine that instead of a monkey typing words, we have an algorithm generating
!
the set of all peptides scoring at least threshold against a spectral vector Spectrum. We
will henceforth call this set of high-scoring peptides a spectral dictionary, denoted
!
D ICTIONARYthreshold (Spectrum) .
!
For a PSM (Peptide, Spectrum), we will use the term PSM dictionary, denoted
!
D ICTIONARY (Peptide, Spectrum), to refer to the spectral dictionary
255
CHAPTER 11
!
D ICTIONARY S CORE(Peptide,Spectrum
! (Spectrum) .
)
!
For DinosaurPSM, the PSM dictionary is D ICTIONARY 19 (DinosaurSpectrum).
Instead of checking which words generated by the monkey occur in an English
dictionary, we will match peptides from a spectral dictionary against a proteome. If we
find matches, then we must decide whether these matches represent biologically valid
PSMs or just statistical artifacts. To make this decision, we consider
!
E ( D ICTIONARYthreshold (Spectrum), n) ,
the expected number of peptides in a decoy proteome of length n that would occur
!
in D ICTIONARYthreshold (Spectrum). If this number is larger than 1, then there is nothing
!
surprising in finding a peptide that scores threshold against Spectrum. We have therefore
formulated our statistical significance test as a special case of the Monkey and the
Typewriter Problem.
Expected Number of High-Scoring Peptides Problem:

Find the expected number of high-scoring peptides against a given spectrum in a decoy
proteome.
!
Input: A spectral vector Spectrum and integers threshold and n.
!
Output: E ( D ICTIONARYthreshold (Spectrum), n).
To solve this problem, we will begin with a spectral dictionary consisting of a single
amino acid string Peptide, which we will attempt to match against a randomly generated
string DecoyProteome of length n. Because DecoyProteome was randomly generated, the
probability that Peptide matches the string beginning at a given position of DecoyProteome
is 1/20|Peptide| . We call this expression the probability of Peptide. Therefore, the expected
number of times that Peptide occurs in DecoyProteome is
n |Peptide| + 1 1
⇡ n · |Peptide| .
20|Peptide| 20
Next, assume that a set of peptides Dictionary contains multiple amino acid strings
of arbitrary lengths. Using the above approximation, the expected number of matches
between strings in Dictionary and DecoyProteome can be approximated as
256
!
1
E (Dictionary, n) ⇡ n · Â 20|Peptide|
.
each peptide Peptide in Dictionary
We will refer to the sum inside the parentheses above as the probability of Dictionary,
denoted Pr(Dictionary), so that the preceding approximation can be written as
E (Dictionary, n) ⇡ n · Pr(Dictionary) .
!
We have thus reduced the statistical analysis of a PSM (Peptide, Spectrum) to the compu-
!
tation of Pr( D ICTIONARY (Peptide, Spectrum)), the probability of the PSM dictionary.
!
STOP and Think: Can Pr( D ICTIONARY (Peptide, Spectrum)) be greater than 1?
You may be wondering why we have used the probabilistic notation Pr(Dictionary). To
!
learn why Pr( D ICTIONARY (Peptide, Spectrum)) is indeed a probability (and thus cannot
exceed 1), see DETOUR: The Probabilistic Space of Peptides in a Spectral Dictionary. PAGE 278
Probability of Spectral Dictionary Problem:

Find the probability of a spectral dictionary for a given spectrum and score threshold.
!
Input: A spectral vector Spectrum and an integer threshold.
!
Output: The probability of D ICTIONARYthreshold (Spectrum).
It seems that we are finally ready to test the statistical significance of DinosaurPSM.
We simply need to first construct the PSM dictionary D ICTIONARY (DinosaurPSM) and
then compute n · Pr( D ICTIONARY (DinosaurPSM)), where n is the length of the string
formed by concatenating the UniProt+ database. If this value is small (say, 0.001), then
we will be able to argue that DinosaurPeptide is a T. rex peptide rather than a statistical
artifact.
Unfortunately, D ICTIONARY (DinosaurPSM) contains over 200 billion peptides, and
generating it would be extremely time-consuming. Can we somehow compute the
probability of this dictionary without having to generate it?
257
CHAPTER 11
Spectral Dictionaries
We will first compute the number of peptides in a spectral dictionary, since this simpler
problem will provide insights on how to compute the probability of a spectral dictionary.
Size of Spectral Dictionary Problem:

Find the size of the spectral dictionary for a given spectral vector and score threshold.
!
Input: A spectral vector Spectrum and an integer threshold.
!
Output: The number of peptides in D ICTIONARYthreshold (Spectrum).
We will use dynamic programming to solve the Size of Spectral Dictionary Problem.
!
Given a spectral vector Spectrum = (s1 , . . . , sm ), we define its i-prefix (for i between 1
!
and m) as Spectrumi = (s1 , . . . , si ) and introduce a variable S IZE (i, t) as the number of
!
peptides Peptide of mass i such that S CORE (Peptide, Spectrumi ) is equal to t. For example,
!
consider the spectral vector Spectrum = (4, -3, -2, 3, 3, -4, 5, -3, -1, -1, 3, 4, 1, 3) of length
14 and the toy amino acid alphabet consisting of amino acids X and Z with respective
masses 4 and 5. There are only three peptides of mass 13 (XXZ, XZX, and ZXX); the
!
first two peptides have score 1 against Spectrum13 , and the third has score 3. Thus,
S IZE (13, 1) = 2, S IZE (13, 3) = 1, and S IZE (13, t) = 0 for all values of t other than 1 and
3.
The key to establishing a recurrence relation for computing S IZE (i, t) is to realize
that the set of peptides contributing to S IZE (i, t) can be split into 20 subsets depending
on their final amino acid a. Each peptide ending in a specific amino acid a results in a
shorter peptide with mass i | a| and score t si if we remove a from the peptide (here,
| a| denotes the mass of a). Thus, as illustrated in Figure 11.12,
S IZE (i, t) = Â S IZE (i | a |, t si ) .

all amino acids a
Since there is a single “empty” peptide of length zero, we initialize S IZE (0, 0) = 1.
We also define S IZE (0, t) = 0 for all possible scores t, and set S IZE (i, t) = 0 for negative
values of i. Using the above recurrence, we can compute the size of a spectral dictionary
!
of Spectrum = (s1 , . . . , sm ) as
!
11H D ICTIONARYthreshold (Spectrum) = Â S IZE (m, t) .
t threshold
258
0 4 -3 -2 3 3 -4 5 -3 -1 -1 3 4 1 3
t\i 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
0 1 0 0 0 0 0
1 0 0 0 0 0 0
2 0 0 0 0 0 0
3 0 0 0 0 1 1
4 0 0 0 0 0 0
5 0 0 0 0 0 0
0 4 -3 -2 3 3 -4 5 -3 -1 -1 3 4 1 3
t\i 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
0 1 0 0 0 0 0 0 0 1 0
1 0 0 0 0 0 0 0 0 0 0
2 0 0 0 0 0 0 0 0 0 2
3 0 0 0 0 1 1 0 0 0 0
4 0 0 0 0 0 0 0 0 0 0
5 0 0 0 0 0 0 0 0 0 0
0 4 -3 -2 3 3 -4 5 -3 -1 -1 3 4 1 3
t\i 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
0 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
2 0 0 0 0 0 0 0 0 0 2 1 0 0 0 0
3 0 0 0 0 1 1 0 0 0 0 0 0 0 2 0
4 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0
5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3
FIGURE 11.12 Computing SIZE (i, t) for the toy alphabet consisting of the amino acids
X and Z with respective masses 4 and 5 and the spectral vector (4, -3, -2, 3, 3, -4, 5, -3,
-1, -1, 3, 4, 1, 3). The boldfaced black entries in the dynamic programming matrices are
computed by summing blue and red entries according to the formula SIZE (i, t) =
SIZE (i 4, t si ) + SIZE (i 5, t si ). For example, in the matrix at the bottom,
SIZE (14, 5) = SIZE (14 4, 5 3) + SIZE (14 5, 5 3) = SIZE (10, 2) + SIZE (9, 2).
259
CHAPTER 11
STOP and Think: The above formula requires computing S IZE (m, t) for all
!
values of t greater than threshold. Given a spectral vector Spectrum, can you find
a value T such that S IZE (m, t) is equal to zero when t > T?
EXERCISE BREAK: Compute | D ICTIONARY (DinosaurPSM)|.
Note that the equation for the probability of a dictionary,
1
Pr(Dictionary) = Â 20|Peptide|
,
is similar to an equation for the size of a dictionary,
|Dictionary| = Â 1.
This similarity suggests that we can derive a recurrence for the probability of a dictionary
using arguments similar to those used to find the size of a dictionary.
Define Pr(i, t) as the sum of probabilities of all peptides with mass i for which
! !
S CORE (Peptide, Spectrumi ) is equal to t. The set of peptides contributing to Pr(i, t) can
be split into 20 subsets depending on their final amino acid. Each peptide Peptide ending
in a specific amino acid a results in a shorter peptide Peptidea if we remove a; Peptidea
has mass i | a| and score t si . Since the probability of Peptide is 20 times smaller than
the probability of Peptidea , the contribution of Peptide to Pr(i, t) is 20 times smaller than
contribution of Peptidea to Pr(i | a|, t si ). Therefore, Pr(i, t) can be computed as
1
Pr(i, t) = Â 20
· Pr(i | a |, t si ) ,
all amino acids a
which differs from the recurrence for computing S IZE (i, t) only in the presence of the
factor 1/20.
We can now compute the probability of a spectral dictionary as
!
11I Pr( D ICTIONARYthreshold (Spectrum)) = Â Pr(m, t) .
t threshold
In particular, we find that D ICTIONARY (DinosaurPSM) consists of 219,136,251,374 pep-

tides and has probability 0.00018. We are therefore ready to test the statistical signifi-
cance of DinosaurPSM found in searches against the UniProt+ database of length n =
194,613,142 (comprising 546,799 proteins).
260
Our aim is to compute n · Pr( D ICTIONARY (DinosaurPSM)), as it approximates the

number of peptides from D ICTIONARY (DinosaurPSM) that we expect to find in a decoy
proteome of length n. Since Pr( D ICTIONARY (DinosaurPSM)) = 0.00018, we have that
n · Pr( D ICTIONARY (DinosaurPSM)) = 35,311 .
We therefore expect to find tens of thousands of peptides scoring at least as high as

!
DinosaurPeptide (against DinosaurSpectrum) in a decoy database, and so there is nothing
surprising about finding DinosaurPSM while searching the UniProt+ database! We
therefore conclude that DinosaurPeptide is a statistical artifact rather than a real T. rex
peptide. But what about the other T. rex peptides?
EXERCISE BREAK: Compute probabilities of spectral dictionaries for all other

T. rex PSMs reported by Asara. Are these PSMs statistically significant?
T. rex Peptides: Contaminants or Treasure Trove of Ancient Proteins?
The hemoglobin riddle
Upon receiving criticism regarding the statistical foundations of his claims, Asara
acknowledged some of the problems with his analysis, withdrew DinosaurPeptide as
an explanation for DinosaurSpectrum, changed some of his previously proposed T.
rex peptides, and released all 31,372 spectra from the T. rex fossil. Afterwards, other
scientists re-analyzed all spectra and verified that although some of the originally
reported T. rex PSMs are questionable, others are statistically solid (Figure 11.13).
However, Asara’s release of T. rex spectra raised more questions than it answered.
In these spectra, Matthew Fitzgibbon and Martin McIntosh identified an additional
spectrum (Figure 11.14) that perfectly matched ostrich hemoglobin, thus adding another
T. rex peptide to the seven collagen peptides in Figure 11.13. The hemoglobin PSM,
which was missed by Asara, is an order of magnitude more statistically significant than
any previously reported T. rex collagen peptide!
It would be shocking if the hemoglobin peptide indeed belonged to T. rex because
hemoglobins are much less conserved than collagens. For example, human beta chain
hemoglobin is 146 amino acids long and has 27, 38, and 45 amino acid differences with
mouse, kangaroo, and, chicken respectively. Furthermore, intact hemoglobin peptides
have never been found in much younger and widely available fossils, such as the bones
of extinct cave bears. These fossils are so common in European caves that they were
261
CHAPTER 11
used as a source of phosphates to produce gunpowder during World War I.
ID Peptide Protein Probability n · Probability

P1 GLVGAPGLRGLPGK Collagen a1t2 1.8 · 10-4 36,000
P2 GVVGLPoh GQR Collagen a1t1 7.6 · 10-8 16
P3 GVQGPPoh GPQGPR Collagen a1t1 7.9 · 10-11 1.6 · 10-2
P4 GATGAPoh GIAGAPoh GFPohGAR Collagen a1t1 3.2 · 10-12 6.4 · 10-4
P5 GLPGESGAVGPAGPIGSR Collagen a2t1 9.9 · 10-14 2.0 · 10-5
P6 GSAGPPoh GATGFPohGAAGR Collagen a1t1 3.2 · 10-14 6.4 · 10-6
P7 GAPGPQGPSGAPoh GPK Collagen a1t1 7.0 · 10-16 1.4 · 10-7
P8 VNVADCGAEALAR Hemoglobin b 7.8 · 10-17 1.6 · 10-8
FIGURE 11.13 The seven candidate T. rex collagen peptides (P1 - P7) reported by Asara
as well as a hemoglobin peptide (P8). The last column shows the probabilities of
the PSM dictionaries formed by these peptides. Red symbols indicate mutated amino
acids compared to peptides in the UniProt database. The amino acid Poh stands for
hydroxyproline, a modified form of proline that is common in collagens.
FIGURE 11.14 A high-quality T. rex spectrum matching the ostrich hemoglobin peptide
VNVADCGGAEAIAR. Nearly all possible prefixes and suffixes are represented by high-
intensity peaks; in fact, applying de novo sequencing to this spectrum results in the
same peptide.
Because Asara had analyzed ostrich samples before analyzing the T. rex sample,
Fitzgibbon and McIntosh argued that the hemoglobin peptide might indicate a contami-
nated sample in the form of a carryover, or the identification of leftover peptides hiding
inside a mass spectrometer after a previous experiment. Contamination is a fact of life
in every proteomics laboratory, as mass spectrometrists are never surprised when they
identify human keratin in their samples: the air in any room typically contains millions
of tiny human skin particles.
262
If the hemoglobin peptide is a carryover, then the entire T. rex sample has been
contaminated, implying that all other T. rex peptides should be discarded. However,
Asara maintained that there was no contamination in his experiment and that the ostrich
hemoglobin must be a T. rex peptide, expanding the class of proteins that can survive
for 68 million years beyond just collagens.
Yet if Horner’s T. rex fossil is indeed a treasure trove of ancient proteins, and we
believe that the hemoglobin peptide came from T. rex, then why should we limit our
search to collagen peptides and their mutated variants? Why not search against all
known proteins from all vertebrates? Of course, we should use criteria that are similar
to ones that Asara used, such as allowing for up to one mutation. If we follow this
criterion, then we should supplement Figure 11.13 with a surprisingly diverse set of
peptides from ostrich, chicken, mouse, and human; some of these peptides are shown
in Figure 11.15.
In light of these new peptides, Asara’s claim about finding molecular evidence of
the link between birds and dinosaurs becomes even weaker (for more on the debate
surrounding the claim that birds evolved from dinosaurs, see DETOUR: Are Terres- PAGE 279
trial Dinosaurs Really the Ancestors of Birds?). If we were to attempt to dismiss the
peptides in Figure 11.15 as statistical artifacts, then we might have to throw out the T.
rex peptides in Figure 11.13 as well.
ID Peptide Protein Probability n · Probability

P9 EDCLSGAKPK ATG7 (Chicken) 3.2 · 10-12 6.4 · 10-4
P10 ENAGEDPGLAR DCD (Human) 2.7 · 10-12 5.4 · 10-4
P11 EGVDAGAAGDPER TTL11 (Mouse) 1.2 · 10-12 2.4 · 10-4
P12 SWIHVALVTGGNK CBR1 (Human) 1.2 · 10-12 2.4 · 10-4
P13 SSNVLSGSTLR MAMD1 (Human) 5.9 · 10-13 1.8 · 10-4
P14 DEVTPAYVVVAR ASPM (Mouse) 1.9 · 10-13 3.8 · 10-5
P15 RNVADCGAEALAR HBB (Ostrich) 3.5 · 10-15 7.0 · 10-7
FIGURE 11.15 Matching T. rex spectra against all vertebrate proteins in the UniProt
database (allowing for up to 1 mutation) reveals a diverse set of peptides. Red symbols
indicate mutated amino acids. Note the presence of another ostrich hemoglobin peptide
(P15), which is slightly heavier (by 57 daltons) than the previously reported hemoglobin
peptide in Figure 11.13 (P8). This change in mass may represent either a mutation of V
into R (as shown above) or a modification of an amino acid.
263
CHAPTER 11
The dinosaur DNA controversy
As the “T. rex peptides” paper continues to age, there is no end in sight to its controversy.
Yet it was not the first paper to report the retrieval of genetic material from dinosaurs.
In 1994, Scott Woodward announced that he had sequenced DNA from an 80 million
year-old dinosaur bone. The most vehement critic of his finding was — believe it or not
— Mary Schweitzer, who proved that Woodward had only sequenced contaminated
human DNA.
The moral is that although we often present scientific discoveries as clear and incon-
trovertible, the reality is that some of the interesting avenues of modern science often
fall short of this ideal. In a sense, the academic battleground is part of the appeal of
becoming a scientist in the first place. But we also cannot help but wonder if we would
have a conclusive answer to whether Horner’s fossil really contained dinosaur peptides
if it had originally been shared with dozens of independent researchers, who would
have undoubtedly unearthed the shocking appearance of hemoglobin in the T. rex
samples. Fittingly, in their criticism of Woodward’s “dinosaur DNA” paper, Schweitzer
wrote, “real advance in [paleontology] will come only when it is demonstrated that
those studies can be replicated in independent laboratories.”
Epilogue: From Unmodified to Modified Peptides
Post-translational modifications
The PSMS EARCH algorithm can only identify a peptide if it occurs in a proteome
without mutations. Yet some of the peptides in Figure 11.13 are mutated.
STOP and Think: How could we generalize PSMS EARCH to find mutated
peptides?
To find a highest-scoring peptide with up to k mutations matching a spectral vector,

we could generate all mutated variants of all peptides in the proteome, concatenate
them into an amino acid string MutatedProteome, and then run PSMS EARCH on Mutat-
edProteome. Unfortunately, the number of mutated peptides will be so large that it may
render PSMS EARCH impractical, even if we allow at most one mutation per peptide.
STOP and Think: How many mutated peptides with at most k mutations are
there for a given peptide of length n?
264
In addition to searching for mutated peptides, we will also need to search for post-
translational modifications, which alter amino acids after a protein has been translated
from RNA. In fact, most proteins are modified after translation, and hundreds of types
of modifications have been discovered. For example, the enzymatic activity of many
proteins is regulated by the addition or removal of a phosphate group at a specific
amino acid (Figure 11.16). This process, called phosphorylation, is reversible; protein
kinases add phosphate groups, whereas protein phosphatases remove them.
O R O R
H 2N H 2N
N N
H H
CH2 O CH2 O
OH OH
O
P
O O
FIGURE 11.16 Tyrosine (left) and its post-translational modification into phosphorylated
tyrosine (right).
In fact, you may have noticed in Figure 11.13 that most candidate T. rex peptides
have a modification transforming proline (mass 97) into hydroxyproline (mass 113).
Hydroxyproline is a major component of collagens that is important to collagen stability
and that comprises roughly 4% of all amino acids in humans.
There are also important but rare post-translational modifications such as diph-
thamide. This modification of histidine only appears in a single protein (protein
synthesis elongation factor-2), but it is universal across all eukaryotes! Researchers
showed that diphthamide is the target for several toxins secreted by various pathogenic
bacteria, which raises the question of why all eukaryotes would retain this modification
if it makes them so vulnerable to pathogens — it must serve some important but still
unknown function in normal physiology.
Searching for modifications as an alignment problem
A modification of mass d applied to an amino acid results in adding d to the mass of

this amino acid. For example, d = 80 for phosphorylated amino acids (serine, threonine,
and tyrosine), d = 16 for the modification of proline into hydroxyproline, and d = 1
265
CHAPTER 11
for the modification of lysine into allysin. If d is positive, then the resulting modified
!
peptide has a peptide vector that differs from the original peptide vector Peptide by
!
inserting a block of d zeroes before the i-th occurrence of 1 in Peptide. In the more rare
case that d is negative, the modified peptide corresponds to deleting a block of |d| zeroes
!
from Peptide (Figure 11.17).
We will use the term block indel to refer to the addition or removal of a block of
consecutive zeroes from a binary vector. Thus, applying k modifications to an amino
!
acid string Peptide corresponds to applying k block indels to its peptide vector Peptide.
We define VARIANTS k (Peptide) as the set of all modified variants of Peptide with up to k
modifications.
!
Given a peptide Peptide and a spectral vector Spectrum, our goal is to find a modified
!
peptide from VARIANTS k (Peptide) with maximum score against Spectrum.
Spectral Alignment Problem:

Given a peptide and a spectral vector, find a modified variant of this peptide that maximizes
the peptide-spectrum score, among all variants of the peptide with up to k modifications.
!
Input: An amino acid string Peptide, a spectral vector Spectrum, and an
integer k.
!
Output: A peptide of maximum score against Spectrum among all peptides
in VARIANTS k (Peptide).
A brute force approach to the Spectral Alignment Problem would score each peptide
in VARIANTS k (Peptide) against the spectrum. We need to solve this problem more effi-
ciently because our more ambitious goal is to solve the following problem, which will
require multiple applications of an algorithm solving the Spectral Alignment Problem.
Modification Search Problem:

Given a spectrum and a proteome, find a peptide of maximum score against this spectrum
among all modified peptides in the proteome with up to k modifications.
!
Input: A spectral vector Spectrum, an amino acid string Proteome and an
integer k.
! !
Output: A peptide Peptide that maximizes S CORE(Peptide, Spectrum) among
all modified variants of peptides from Proteome with up to k modifications.
266
Like modifications, mutations can also be viewed as block indels; for example, the muta-
tion of V (integer mass 99) into R (integer mass 156) in peptide P9 from Figure 11.15 can
be viewed as a block insertion with d = 156 99 = 57. We can therefore transform the
Modification Search Problem into a “Mutation Search Problem” by simply substituting
the word “modification” by “mutation” in the problem statement. In this new problem,
a mutation of mass d applied to an amino acid corresponds to the difference between
the mass of this amino acid and another one. For example, mutations of valine (integer
mass 99) correspond to modifications with the integer masses -42, -28, -12, -2, 2, 4, 14,
15, 16, 29, 30, 32, 38, 48, 57, 64, and 87.
XZZXX 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 1 0 0 0 1
XZ+3 ZXX 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 0 1 0 0 0 1
XZ+3 ZX-2 X 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 1 0 1 0 0 0 1
FIGURE 11.17 Transforming the peptide XZZXX into the peptide XZ+3 ZXX corresponds
to inserting a block of three zeroes (shown in red) before the second occurrence of 1
in the peptide vector of XZZXX. Transforming the peptide XZ+3 ZXX into the peptide
XZ+3 ZX-2 X corresponds to deleting a block of two zeroes (shown in green) before the
fourth occurrence of 1 in the peptide vector of XZ+3 ZXX.
Building a Manhattan grid for spectral alignment
The problem of finding a highest scoring peptide vector having up to k block indels
recalls sequence alignment problems. This insight suggests that we should frame the
Spectral Alignment Problem as an instance of the Longest Path in a DAG Problem.
STOP and Think: How would you build a DAG to solve this problem?
Consider the (m + 1) ⇥ (m + D + 1) Manhattan grid in which every node (i, j) is con-

nected to every node (i0 , j0 ) for 0  i < i0  m and 0  j < j0  m + D (Figure 11.18).
We call this graph S OUTHEAST (m, m + D) and call the nodes (0, 0) and (m, m + D) source
and sink, respectively.
Next, consider an amino acid string Peptide = a1 . . . an of mass m and its modified
variant Peptidemod = a10 . . . a0n of mass m + D. Define the path PATH (Peptide, Peptidemod )
267
CHAPTER 11
in S OUTHEAST (m, m + D) consisting of n edges:
(0, 0) ! ( M ASS ( a1 ), M ASS ( a10 ))

! ( M ASS ( a1 a2 ), M ASS ( a10 a20 ))
! ...
! ( M ASS ( a1 . . . an ), M ASS ( a10 . . . a0n ))
= (m, m + D) .
0 1 2 3
0
2
FIGURE 11.18 The graph SOUTHEAST (2, 3). Each node in the graph is connected to
every node lying to its south and east, with the exception of nodes in the same row
and column.
For example, consider the toy amino acid alphabet containing X, Y, and Z with respec-
tive masses 2, 3, and 4. The blue path in Figure 11.19 indicates PATH (XYYZX, XY+2 YZ-3 X):
(0, 0) ! (2, 2) ! (5, 7) ! (8, 10) ! (12, 11) ! (14, 13).
Except for the initial node (0, 0), every node (i, j) on this path indicates that the i-th
! !
element of Peptide = (0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 0, 1, 0, 1) and the j-th element of Peptidemod
= (0, 1, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 1) are both 1.
An edge connecting (i, j) to (i0 , j0 ) in S OUTHEAST (m, m + D) is called diagonal if
i i = j0 j and non-diagonal otherwise. Note that if an amino acid of Peptidemod is un-
0
modified, then the edge corresponding to this amino acid in PATH (Peptide, Peptidemod )
is diagonal. An amino acid a with modification d in Peptidemod corresponds to a non-
diagonal edge in this path connecting some node (i, j) with the node (i + | a|, j + | a| + d).
We are now ready to solve the Spectral Alignment Problem for an amino acid string
!
Peptide = a1 . . . an of mass m and a spectral vector Spectrum = s1 . . . sm+D . We know
that the modified variant of Peptide solving this problem must have mass m + D. We
can thus represent all modified peptides of mass m + D in VARIANTS k (Peptide) as paths
268
in S OUTHEAST (m, m + D) from source to sink with at most k non-diagonal edges (Fig-
!
ure 11.19 (top)). We refer to these modified peptides as VARIANTS k (Peptide, Spectrum).
STOP and Think: Does every path from source to sink in S OUTHEAST(m, m + D)
correspond to a candidate modified peptide of mass m + D?
!
Although every peptide in VARIANTS k (Peptide, Spectrum) corresponds to a path from
source to sink in S OUTHEAST(m, m + D), many paths in this graph do not correspond
to such modified peptides. Indeed, since Peptide is fixed, any path corresponding to a
modified variant of Peptide will only pass through rows with indices
0, M ASS ( a1 ), M ASS ( a1 a2 ), . . . , M ASS ( a1 . . . an ) = m ,
shown as rows with darker nodes in Figure 11.19 (top).

Thus, nodes in other rows can be safely removed from S OUTHEAST(m, m + D),
!
which results in the PSM graph, denoted PSMG RAPH (Peptide, Spectrum), and shown
in Figure 11.19 (bottom). Note that the n + 1 rows of nodes in the PSM graph have
indices i equal to
0, M ASS ( a1 ), M ASS ( a1 a2 ), . . . , M ASS ( a1 . . . an )
rather than the indices 0, 1, . . . , n.

In the PSM graph, all edges entering into the row with index i = M ASS ( a1 . . . at )
originate at the row with index M ASS ( a1 . . . at 1 ). We therefore define D IFF (i ) as the
mass of the amino acid at . For the peptide XYYZX, D IFF(2) = M ASS(X) = 2, D IFF(5) =
M ASS(Y) = 3, D IFF(8) = M ASS(Y) = 3, D IFF(12) = M ASS(Z) = 4, and D IFF(14) = M ASS(X)
= 2.
STOP and Think: Can you assign weights to the nodes of

!
PSMG RAPH (Peptide, Spectrum) so that the total weight of a path from
source to sink corresponding to a modified peptide Peptidemod is equal to
!
S CORE (Peptidemod , Spectrum)?
!
Given a spectral vector Spectrum = (s1 , . . . , sm+D ), we assign a weight of s j to every node
(i, j) in column j of the PSM graph. With this assignment of weights, the total weight of
!
nodes in the path corresponding to Peptidemod is equal to S CORE (Peptidemod , Spectrum).
Solving the Spectral Alignment Problem is therefore equivalent to finding a path in
the PSM graph with maximum total node weight among all paths with at most k
non-diagonal edges.
269
CHAPTER 11
Peptidemod
X Y+2 Y Z-3 X
X-1 Y Y-2 Z X+2
0 1 2 3 4 5 6 7 8 9 10 11 12 13
0
X 2
Y 5
Peptide 7
Y 8
10
11
Z 12
13
X 14
Peptidemod
X Y+2 Y Z-3 X
X-1 Y Y-2 Z X+2
0 1 2 3 4 5 6 7 8 9 10 11 12 13
0
X 2
Y 5
Peptide
Y 8
Z 12
X 14
FIGURE 11.19 (Top) Two paths in SOUTHEAST (14, 13) formed by the modifications
XY+2 YZ-3 X and X-1 YY-2 ZX+2 of XYYZX, where X, Y, and Z have respective masses
2, 3, and 4. Diagonal edges are solid, whereas non-diagonal edges are dashed. The
indices of darker rows in the graph correspond to occurrences of 1 in the peptide vector
of XYYZX. The darker nodes in this graph form the PSM graph. (Bottom) Another
representation of the PSM graph obtained by removing the light nodes from the graph
above.
270
Spectral alignment algorithm
We already know how to find a longest path in a node-weighted DAG. However, it is

not clear how to find a longest path in a DAG under the additional constraint that this
path has at most k non-diagonal edges.
As a workaround, we will convert the two-dimensional PSM graph from Figure 11.19
(bottom) into a three-dimensional spectral alignment graph consisting of k + 1 layers
(Figure 11.20), where the node set in each layer coincides with the node set of the PSM
graph. This graph will have nodes (i, j, t), where 0  i  m, 0  j  m + D, and
0  t  k. Each such node inherits the weight of the node (i, j) in the PSM graph (i.e.,
!
the amplitude s j from Spectrum = (s1 , . . . , sm+D )).
As for edges, each of the k + 1 layers of the spectral alignment graph will inherit all
diagonal edges from the PSM graph, i.e., each diagonal edge from (i, j) to (i + x, j + x ) in
the PSM graph will correspond to the k + 1 edges (i, j, t) to (i + x, j + x, t) for 0  t  k.
STOP and Think: The layers in the constructed graph are now disconnected.
How should we connect them?
For each non-diagonal edge connecting (i, j) to (i0 , j0 ) in the PSM graph, we will generate
k edges connecting consecutive layers in the spectral alignment graph by connecting
(i, j, t) to (i0 , j0 , t + 1) for all t between 0 and k 1. Every path in the spectral alignment
graph from (0, 0, 0) to (m, m + D, t) corresponds to a modified version of peptide with t
modifications (Figure 11.20). The zero-th layer of this graph will store scores of peptides
with no modifications, the first layer will store scores of peptides with one modification,
and so on.
To solve the Spectral Alignment Problem, we will define S CORE (i, j, t) as the maxi-
mum score of all paths connecting node (0, 0, 0) to node (i, j, t) in the spectral alignment
graph. Note that this score is equal to the weight s j assigned to node (i, j, t) plus the
maximum of the scores of all predecessors of node (i, j, t). One of these predeces-
sors, (i D IFF (i ), j D IFF (i ), t), is located in the same layer, and j other predecessors
((i D IFF (i ), j0 , t 1) for j0 < j) are located in the previous layer. This reasoning results
in the following recurrence for computing S CORE (i, j, t),
(
S CORE (i D IFF (i ), j D IFF (i ), t)
S CORE (i, j, t) = si + max .
j <j
0
S CORE (i D IFF (i ), j0 , t 1)
The maximum score of all peptides with at most k modifications is therefore the maxi-
mum value of S CORE (m, m + D, t) as t ranges from 0 to k. To initialize the recurrence,
we assume that S CORE (0, 0, 0) = 0 and that S CORE (0, 0, t) = • for all 1  t  k.
271
CHAPTER 11
0 1 2 3 4 5 6 7 8 9 10 11 12 13
0
5
Layer 0
8
12
14
5
Layer 1
8
12
14
5
Layer 2
8
12
14
5
Layer 3
8
12
14
FIGURE 11.20 The PSM graph from Figure 11.19 (bottom) transformed into a spectral
alignment graph with four layers (k = 3). The blue and red paths from the PSM graph
in Figure 11.19 (bottom) correspond to the blue and red paths shown, which represent
the respective modified variants XY+2 YZ-3 X and X-1 YY-2 ZX+2 of XYYZX. The blue
path terminates in Layer 2 because it corresponds to a peptide with two modifications,
whereas the red path terminates in Layer 3 because it corresponds to a peptide with
three modifications.
272
Although the above recurrence computes the score of a modified peptide solving
the Spectral Alignment Problem, we also need to reconstruct this peptide. In order to
achieve this goal, we will need to implement a backtracking approach similar to the 11J
backtracking approach described in Chapter 5, which we leave to you as an exercise.
EXERCISE BREAK: What is the running time of the spectral alignment algo-
rithm?
CHALLENGE PROBLEM: In addition to T. rex, Asara also analyzed a 200,000

year-old mastodon fossil. The extinction of the elephant-like mastodons 10,000
years ago was caused by a combination of climate change and hunting by humans
armed with stone weapons. In contrast to dinosaurs, it is not surprising that
scientists routinely identify peptides from recently extinct species like mastodons
or cave bears.
Analyze the mastodon collagen peptides reported in Asara’s 2007 paper and
decide which of them form statistically significant PSMs. Can you identify other
statistically significant mastodon peptides that this paper missed? Can you
find non-collagen peptides (especially hemoglobin peptides) matching spectra
from mastodons? Can you determine the number of different types of post-
translational modifications of mastodon peptides by solving the Modification
Search Problem?
Detours
Gene prediction
To predict split genes, researchers often attempt to recognize the locations of splicing
signals at exon-intron junctions. For a simple example, the dinucleotides AG and GT
on either side of an exon are highly conserved (Figure 11.21). To improve the accuracy
of this approach (known as statistical gene prediction), researchers look for genomic
features appearing frequently in exons and infrequently in introns.
Attempts to improve the accuracy of statistical gene prediction methods have led
to similarity-based gene prediction approaches, which are based on the observation
that a newly sequenced gene is often similar to a known gene in another species. For
example, 99% of mouse genes have human analogs.
However, we cannot simply look for a similar sequence in the mouse genome based
on known human genes, since the exon sequence and partition of a gene into exons
273
CHAPTER 11
in different species may be different. To address this complication, similarity-based

approaches sometimes look for a set of putative exons in the mouse genome whose
concatenation fits a known human protein.
GT AG GT AG
FIGURE 11.21 A split gene with exons (red) separated by introns (blue). Introns typically
start with GT and end with AG.
Yet not all genes are split genes. In fact, bacteria do not have split genes at all, which
simplifies bacterial gene prediction. Such genes begin with a start codon that codes for
methionine (typically ATG but also sometimes GTG or TTG) and end with a stop codon
(TAA, TAG, or TGA).
We can represent a genome of length n as a sequence of n/3 codons. The stop codons
break this sequence into segments between every pair of consecutive stop codons. The
suffixes of these segments that begin at the first start codon within a segment are called
open reading frames, or ORFs. ORFs within a single genome may overlap because
there are six possible reading frames.
In a randomly generated DNA string, we would expect to find a stop codon ap-
proximately every 64 nucleotides within a given reading frame. Yet the typical length
of a bacterial gene is on the order of 1,000 nucleotides. A gene prediction algorithm
might therefore select ORFs longer than some threshold length as candidate genes.
Unfortunately, such an algorithm would also fail to detect short genes.
Many gene prediction algorithms also rely on subtle statistical differences between
coding and non-coding regions, such as biases in codon usage, or the frequency of each
codon. For example, there are six codons encoding leucine, but whereas CUG encodes
47% of all occurrences of leucine in E. coli, CUA encodes only 4%. Therefore, an ORF
with many more occurrences of CUG than CUA is a candidate gene.
Bacterial gene prediction also takes advantage of several conserved motifs often
found in genomic regions near start of RNA transcription. For example, the Pribnow
box is a six-nucleotide sequence with consensus TATAAT that is an essential component
for initiating transcription in bacteria.
Finding all paths in a graph
In the 19th Century, Charles Pierre Trémaux developed an algorithm for navigating
mazes, described as follows. As you walk through a maze, drag a piece of chalk along
274
the ground behind you. When you reach a junction, unmarked paths correspond to
yet unexplored paths. So take an unexplored path for as long as you can, until you
encounter a dead-end or a junction at which all the outgoing paths are marked. In this
case, backtrack your steps until you encounter the exit or a junction with an unmarked
path, or until you arrive at your starting point, in which case the maze has no solution.
Trémaux’s maze algorithm is an example of depth-first search (DFS), a technique
for traversing the nodes of a graph. DFS starts at a given node and explores the graph as
far as possible until it reaches a node that has no outgoing edges or from which we have
already explored all edges. We then backtrack until we reach a node with unexplored
edges. The preorder traversal that we encountered in DETOUR: From Suffix Trees to PAGE 171
Suffix Arrays offers one example of a DFS applied to rooted trees.
The following recursive algorithm offers a DFS-inspired approach to finding all
paths between a node v and a node sink in a DAG Graph.
ALLPATHS(Graph, v, sink)
if v = sink
Paths the set of paths consisting of the single-node path v
else
Paths an empty set of paths
for all outgoing edges (v, w) from v
PathsFromDescendant ALLPATHS (Graph, v, sink)
add (v, w) as the first edge to each path in PathsFromDescendant
add PathsFromDescendant to Paths
return Paths
The Anti-Symmetric Path Problem
In the main text, we saw that not every path from source to sink in G RAPH (Spectrum)
represents a solution of the Decoding an Ideal Spectrum Problem. This issue is caused
by the fact that every mass in the spectrum may be interpreted as either a mass of a
prefix or a mass of a suffix. Therefore, every node corresponding to mass s has a “twin”
node (corresponding to M ASS (Peptide) s). Given an arbitrary node and its twin in
G RAPH (Spectrum), a correct path from source to sink must pass through exactly one of
these nodes in order to spell out a solution.
The Decoding an Ideal Spectrum Problem is a particular case of the following more
general problem. Given a collection of forbidden pairs of nodes in a graph (when
reconstructing peptides, forbidden pairs correspond to twins), a path in a graph is
275
CHAPTER 11
called anti-symmetric if it contains exactly one node from every forbidden pair.
Anti-Symmetric Path Problem:

Find an anti-symmetric path in a DAG.
Input: A DAG with nodes source and sink and a set of forbidden pairs of
nodes in this DAG.
Output: An anti-symmetric path in this DAG from source to sink.
The Anti-Symmetric Path Problem is NP-Hard, but we should not give up hope of
finding an efficient algorithm for the Decoding an Ideal Spectrum Problem because
the latter is a specific instance of the former. In particular, forbidden pairs in peptide
sequencing have the additional property that the sum of masses from each forbidden
pair is equal to the mass of the entire peptide. In fact, there is a polynomial algorithm
solving the Anti-Symmetric Path Problem for a DAG satisfying this additional property,
but this algorithm is outside the scope of this book.
Transforming spectra into spectral vectors
Our goal is to develop a probabilistic model describing how a peptide vector generates
an integer-valued spectrum and to use this model to transform a spectrum into a spectral
vector. To address this problem, we will first introduce an abstract model that seemingly
has nothing to do with peptide sequencing but rather describes a probabilistic process
that transforms a peptide vector P = ( p1 , . . . , pm ) into a binary vector X = ( x1 , . . . , xm )
of the same length. We will see later how ideas developed for this model help us analyze
real spectra.
We define the probability that P generates X as Pr( X | P) = Pim=1 Pr( xi | pi ), where
Pr( xi | pi ) is the probability that pi in P generates xi in X (Figure 11.22). For example, the
probability that a 1 in P generates a 1 in X is written Pr(1|1) and is equal to some pa-
rameter r. The probability that a 0 in P generates a 1 in X is written Pr(1|0) and is equal
to some parameter q. The probability of a 1 in P generating a 0 in X is Pr(0|1) = 1 r,
and the probability of a 0 in P generating a 0 in X is Pr(0|0) = 1 q.
For the toy amino acid alphabet containing just two amino acids with masses 2 and
3, Figure 11.23 illustrates the peptide vector P = (0, 1, 0, 1, 0, 0, 1) generating the binary
vector X = (0, 0, 0, 1, 1, 0, 1) with probability
276
symbol in P
0 1
0 1 j 1 r
symbol in X
1 j r
FIGURE 11.22 A matrix describing a probabilistic process that transforms a peptide
vector P into a binary vector X.
Pr( X | P) = (1 q ) · (1 r ) · (1 q ) · r · q · (1 q ) · r.
We are interested in the following problem.
Most Likely Peptide Vector Problem:

Find a most likely peptide vector for a given binary vector.
Input: A binary vector X and parameters r and q such that 0  r, q  1.

Output: A peptide vector P that maximizes Pr(X|P) as defined by probabili-
ties r and q among all possible peptide vectors.
The solution of this problem is given in DETOUR: Solving the Most Likely Peptide PAGE 280
Vector Problem.
peptide vector P 0 1 0 1 0 0 1
binary vector X 0 0 0 1 1 0 1
Pr(X|P ) Pr(0|0) · Pr(0|1) · Pr(0|0) · Pr(1|1) · Pr(1|0) · Pr(0|0) · Pr(1|1)
(1 - j) · (1 - r) · (1 - j) · r · j · (1 - j) · r
!
Pr(X | 0 ) Pr(0|0) · Pr(0|0) · Pr(0|0) · Pr(1|0) · Pr(1|0) · Pr(0|0) · Pr(1|0)
(1 - j) · (1 - j) · (1 - j) · j · j · (1 - j) · j
Pr(0|1) Pr(1|1) Pr(1|1)
LIKELIHOOD(X|P) 1 · · 1 · · 1 · 1 ·
Pr(0|0) Pr(1|0) Pr(1|0)
1 r r r
1 · · 1 · · 1 · 1 ·
1 j j j
1 r r r
log2 (LIKELIHOOD(X|P)) 0 +log2 + 0 + log2 + 0 + 0 + log2
1 j j j
FIGURE 11.23 A peptide vector P = (0, 1, 0, 1, 0, 0, 1) generates a binary vector X =
(0, 0, 0, 1, 1, 0, 1) with probability Pr(X|P).
277
CHAPTER 11
The infinite monkey theorem
In Jonathan Swift’s Gulliver’s Travels, a professor of the Grand Academy of Legado

asks his students to generate random strings of letters by turning cranks on a machine.
According to the professor, the Academy will eventually crank out brilliant works on
all subjects.
Inspired by Swift’s satirical treatment of certain academics, the infinite monkey
theorem states that an immortal monkey typing an infinite sequence of symbols on a
typewriter will one day reproduce Hamlet. In more technical terms, this theorem states
that an infinite random string contains an arbitrary given text as a substring almost
surely, or with probability equal to 1.
STOP and Think: In 2003, researchers placed a typewriter in a monkey enclosure,

and found that the monkeys typed the letter “S” over and over. It is possible that
an infinite random string generated by a monkey would contain only the letter
“S”. How, then, must this string contain Hamlet almost surely?
The probabilistic space of peptides in a spectral dictionary
In the main text, we defined the probability of Peptide as 1/20|Peptide| , and we defined
the probability of a collection of peptides Dictionary as
1
Pr(Dictionary) = Â 20|Peptide|
.
But why have we used the probabilistic notation? After all, consider the following
exercise, which indicates that it is possible for Pr(Dictionary) to be larger than 1.
EXERCISE BREAK: If Dictionary is the set of all peptides of length at most 10,
what is Pr(Dictionary)?
Yet recall that a peptide can match a spectrum if and only if its mass is equal to the
mass of the spectrum. Thus, no peptide in a spectral dictionary can contain another pep-
tide in the dictionary as a substring, i.e., a spectral dictionary forms a substring-free set.
EXERCISE BREAK: Prove that if Dictionary is a substring-free set, then

Pr(Dictionary)  1.
278
However, you may still be wondering which event corresponds to Pr(Dictionary). More
precisely, what is the “probabilistic space” of underlying outcomes from which Dictio-
nary is formed? The probabilistic space that we propose contains all decoy proteomes of
length n, where n is the length of the longest peptide in a spectral dictionary Dictionary.
In this space, we will assume that each decoy proteome has the same probability. Thus,
our probabilistic space consists of 20n elements, each with probability 1/20n . Note the
change: instead of considering the probabilistic space of all peptides in Dictionary, we
have switched to considering all decoy proteomes.
Each string Peptide in Dictionary appears in exactly 20n |Peptide| decoy proteomes as
their first peptide. The combined probabilities of all these decoy proteomes sum to
1 1
20n |Peptide| · = |Peptide| ,
20n 20
which is Pr(Peptide). Since spectral dictionaries are substring-free, each decoy proteome
has at most one peptide from the spectral dictionary starting at its first position. Thus,
Pr(Dictionary) is simply the combined probability of all decoy proteomes that begin
with one of the peptides in Dictionary.
Are terrestrial dinosaurs really the ancestors of birds?
Aside from the mysterious presence of hemoglobin peptides in Asara’s T. rex spectra,
scientists have recently expressed doubts about the hypothesis that birds evolved from
terrestrial dinosaurs like T. rex and that flight was achieved from the biophysically
improbable ground-up model. This hypothesis assumes that, in order to evolve into
birds, dinosaurs first must have reduced their size while simultaneously developing
feathers (arguably the most complex evolutionary invention for flight).
Most early dinosaur studies favored evidence for a small arboreal animal as the
more logical interpretation for the bird ancestor. This conjecture assumes that before
evolving a system for sustained flight, early birds used gravity-assisted aerodynamics
such as parachuting and gliding (the latter is used by modern flying squirrels). Thus,
the tiny Scansoriopteryx (Figure 11.24), whose fossils contain impressions of feathers and
whose foot adaptations indicate an arboreal lifestyle, competes with T. rex for the honor
of being the ancestor of birds.
If you are interested in learning more about the evolutionary controversy surround-
ing the origin of birds, we suggest two papers, one on either side of the debate:
• “Jurassic archosaur is a non–dinosaurian bird” by Stephen Czerkas and Alan

Feduccia.
279
CHAPTER 11
• “Three crocodilian genomes reveal ancestral patterns of evolution among ar-

chosaurs” by Richard Green et al.
FIGURE 11.24 An artistic recreation of Scansoriopteryx..
Solving the Most Likely Peptide Vector Problem

! !
Define L IKELIHOOD ( X | P) as Pr( X | P)/Pr( X | 0 ), where 0 is an all-zeroes vector con-
sisting of only zeroes. Figure 11.23 illustrates that a peptide vector P = (0, 1, 0, 1, 0, 0, 1)
generates a binary vector X = (0, 0, 0, 1, 1, 0, 1) with
Pr(0|1) Pr(1|1) Pr(1|1)

L IKELIHOOD ( X | P) = · ·
Pr(0|0) Pr(1|0) Pr(1|0)
1 r r r
= · · .
1 q q q
To avoid dealing with extremely small values that result from many multiplications
in L IKELIHOOD(X|P), we will instead use the log-likelihood log2 ( L IKELIHOOD ( X | P)).
Finding a peptide vector maximizing the log-likelihood is equivalent to finding a most
probable peptide vector. Figure 11.23 illustrates that the peptide vector (0, 1, 0, 1, 0, 0, 1)
generates a binary vector (0, 0, 0, 1, 1, 0, 1) with log-likelihood equal to
1 r r r
log2 + log2 + log2 .
1 q q q
We will now transform a binary vector X = ( x1 , . . . , xm ) into a spectral vector S =
(s1 , . . . , sm ) by changing each occurrence of 0 into the amplitude log2 [(1 r)/(1 q )]
and each occurrence of 1 into the amplitude log2 (r/q ). For example, the binary vector
(0, 0, 0, 1, 1, 0, 1) will be transformed into the spectral vector
280
✓ ◆
1 r 1 r 1 r r r 1 r r
log2 , log2 , log2 , log2 , log2 , log2 , log2 .
1 q 1 q 1 q q q 1 q q
Note that log2 ( L IKELIHOOD ( X | P)) is simply the dot product of the peptide vector
P = ( p1 , . . . , pm ) and the spectral vector S = (s1 , . . . , sm ),
P · S = p1 · s1 + · · · + p m · s m .
We denote P · S as S CORE(P, S) (Figure 11.25) and define the score between a peptide
vector and a spectral vector of different length as •. We have therefore transformed
the Most Likely Peptide Vector Problem into the Peptide Sequencing Problem.
peptide vector P 0 1 0 1 0 0 1
binary vector X 0 0 0 1 1 0 1
1 r 1 r 1 r r r 1 r r
spectral vector S log2 log2 log2 log2 log2 log2 log2
1 j 1 j 1 j j j 1 j j
1 r r r
SCORE(P, S) log2 + log2 + log2
1 j j j
FIGURE 11.25 Scoring a peptide vector P against a spectral vector S as the dot product
SCORE (P, S).
Our conversion of a binary vector X into a spectral vector S was based on a simple
probabilistic model (describing how a peptide vector generates a binary vector) with
just two parameters, r and q. To see how spectra are transformed into spectral vectors in
practice, see DETOUR: Selecting Parameters for Transforming Spectra into Spectral PAGE 281
Vectors.
Selecting parameters for transforming spectra into spectral vectors
If mass spectrometers generated binary spectra, then we could start by forming a large
training sample of annotated spectra for which the peptides that generated these spec-
tra are known. We could then estimate r (as the frequency of ones in binary spectra
being generated by ones in peptide vectors) and q (as the frequency of ones in binary
spectra being generated by zeroes in peptide vectors) across all annotated spectra in the
training sample. But since real mass spectrometers generate integer-valued rather than
binary spectra, deriving parameters becomes more complex.
STOP and Think: Can you devise a probabilistic model that would convert the
peptide vector (0, 1, 0, 1, 0, 0, 1) into the integer-valued vector (3, 4, 2, 6, 9, 4, 7)?
281
CHAPTER 11
However, a similar probabilistic model will work if we define the probability of con-
verting zeroes and ones in the peptide vector into various intensities in the real spectra
(rather than into zeroes and ones as before). In fact, the conversion of real spectra into
spectral vectors is based on a similar log-likelihood model that uses dozens of proba-
bilistic parameters. Algorithms for converting real spectra into spectral vectors attempt
to optimize these parameters so that amplitudes at prefix coordinates are maximized
and amplitudes at noise coordinates are minimized.
To derive these parameters, we again need to build a large training sample of
annotated spectra. We can consider all peaks with a certain intensity level in all spectra
and compute which fraction of them are annotated by prefix or suffix peptides. For
example, only 19% and 45% of the ten highest intensity peaks in Collision-Induced
Dissociation spectra (which are similar to ones generated in Asara’s laboratory) are
explained by prefix and suffix peptides, respectively. The remaining high-intensity
peaks are treated as noise. Given a spectrum generated by an unknown peptide vector
P = ( p1 , . . . , pm ), its spectral vector S = (s1 , . . . , sm ) is derived using these frequencies
so that si is the log likelihood ratio log2 (Pr1 /Pr0 ), where Pr1 is an estimate of the
probability that pi = 1, and Pr0 is an estimate of the probability that pi = 0. A complete
discussion of the details of the algorithm for generating spectral vectors is beyond the
scope of this detour.
Figure 11.26 shows the set of prefix masses for DinosaurPeptide along with ampli-
tudes of the spectral vector corresponding to these masses. Note from Figure 11.8
(bottom) that most amplitudes of the spectral vector are negative. The blue elements in
Figure 11.26 correspond to positions that significantly exceed the average amplitude
value.
282
1 2 3 4 5 6 7 8 9 10 11 12 13 14
amino acid G L V G A P G L R G L P G K
mass 57 113 99 57 71 97 57 113 156 57 113 97 57 128
prefix mass 57 170 269 326 397 494 551 664 820 877 990 1087 1144 1272
amplitude -8 +1 -4 -6 -6 +3 +1 -4 -8 +18 +11 -10 -7 0
FIGURE 11.26 Masses of amino acids in GLVGAPGLRGLPGK (second line), prefix
masses for this peptide (third line), and the corresponding elements of the spectral
vector for DinosaurSpectrum, shown in Figure 11.5 (top) (fourth line). Blue elements
correspond to amplitudes that are significantly higher than the average, which is nega-
tive. Note that the second and third tallest peaks in the spectrum (labeled b10 and b11 )
correspond to maximum amplitudes +18 and +11 in the spectral vector of Dinosaur-
Spectrum. Also note that because there is no peak b12 (or y2 ) in Figure 11.5 (bottom),
s1087 = 10 is very small.
Bibliography Notes
T. rex peptides were reported by Asara et al., 2007 and faced criticism in Pevzner, Kim,
and Ng, 2008, Buckley et al., 2008, and Kim et al., 2015. The “dinosaur DNA” paper
by Woodward, Weyand, and Bunnell, 1994 was refuted by Hedges and Schweitzer, 1995.
Czerkas and Feduccia, 2014 recently argued that birds did not evolve from dinosaurs,
while Green et al., 2014 recently argued otherwise.
Chen et al., 2001 solved the Anti-Symmetric Path Problem in the case of graphs
arising from mass spectra. Searching a protein database for the purpose of peptide
identification in mass spectrometry was pioneered by Eng, McCormack, and Yates,
1994. The spectral alignment algorithm was introduced by Pevzner, Dančík, and Tang,
2000. The concept of spectral dictionary and the algorithm for evaluating statistical
significance of PSMs were introduced by Kim, Gupta, and Pevzner, 2008 and Kim et al.,
2009.
283
Bibliography
Aho, A. V. and M. J. Corasick (1975). “Efficient String Matching: An Aid to Bibliographic

Search”. Communications of the ACM Vol. 18: 333–340.
Alon, U., N. Barkai, D. A. Notterman, K. Gish, S. Ybarra, D. Mack, and A. J. Levine (1999).
“Broad patterns of gene expression revealed by clustering analysis of tumor and
normal colon tissues probed by oligonucleotide arrays”. Proceedings of the National
Academy of Sciences Vol. 96: 6745–6750.
Altschul, S. F., W. Gish, W. Miller, E. W. Myers, and D. J. Lipman (1990). “Basic local
alignment search tool.” Journal of Molecular Biology Vol. 215: 403–410.
Arthur, D. and S. Vassilvitskii (2007). “k-means++: The Advantages of Careful Seeding”.
In: Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms.
New Orleans, Louisiana: Society for Industrial and Applied Mathematics, 1027–
1035.
Asara, J. M., M. H. Schweitzer, L. M. Freimark, M. Phillips, and L. C. Cantley (2007).
“Protein Sequences from Mastodon and Tyrannosaurus rex Revealed by Mass Spec-
trometry”. Science Vol. 316: 280–285.
Baldi, P., Y. Chauvin, T. Hunkapiller, and M. A. McClure (1994). “Hidden Markov
models of biological primary sequence information”. Proceedings of the National
Academy of Sciences Vol. 91: 1059–1063.
Bateman, A., E. Birney, L. Cerruti, R. Durbin, L. Etwiller, S. R. Eddy, S. Griffiths-Jones, K.
L. Howe, M. Marshall, and E. L. L. Sonnhammer (2002). “The Pfam Protein Families
Database”. Nucleic Acids Research Vol. 30: 276–280.
285
Baum, L. E., T. Petrie, G. Soules, and N. Weiss (1970). “A Maximization Technique
Occurring in the Statistical Analysis of Probabilistic Functions of Markov Chains”.
The Annals of Mathematical Statistics Vol. 41: pp. 164–171.
Beerenwinkel, N. and M. Drton (2007). “A mutagenetic tree hidden Markov model for
longitudinal clonal HIV sequence data”. Biostatistics Vol. 8: 53–71.
Ben-Dor, A., R. Shamir, and Z. Yakhini (1999). “Clustering Gene Expression Patterns”.
Journal of Computational Biology Vol. 6: 281–297.
Bezdek, J. C. (1981). Pattern Recognition with Fuzzy Objective Function Algorithms. Kluwer
Academic Publishers.
Buckley, M., A. Walker, S. Y. W. Ho, Y. Yang, C. Smith, P. Ashton, J. T. Oates, E. Cappellini,
H. Koon, K. Penkman, B. Elsworth, D. Ashford, C. Solazzo, P. Andrews, J. Strahler,
B. Shapiro, P. Ostrom, H. Gandhi, W. Miller, B. Raney, M. I. Zylber, M. T. P. Gilbert,
R. V. Prigodich, M. Ryan, K. F. Rijsdijk, A. Janoo, and M. J. Collins (2008). “Comment
on ’Protein Sequences from Mastodon and Tyrannosaurus rex Revealed by Mass
Spectrometry’”. Science Vol. 319: 33.
Burrows, D. and M. J. Wheeler (1994). “A block sorting lossless data compression
algorithm”. Technical Report 124, Digital Equipment Corporation.
Cann, R. L., M. Stoneking, and A. C. Wilson (1987). “Mitochondrial DNA and human
evolution.” Nature Vol. 325: 31–36.
Ceppellini, B. R., M. Siniscalco, and C. A. B. Smith (1955). “The Estimation of Gene
Frequencies in a Random-Mating Population”. Annals of Human Genetics Vol. 20:
97–115.
Chen, T., M.-Y. Kao, M. Tepel, J. Rush, and G. M. Church (2001). “A Dynamic Program-
ming Approach to De Novo Peptide Sequencing via Tandem Mass Spectrometry”.
Journal of Computational Biology Vol. 8: 325–337.
Churchill, G. (1989). “Stochastic models for heterogeneous DNA sequences”. Bulletin of
Mathematical Biology Vol. 51: 79–94.
Clayton-Smith, J., J. O’Sullivan, S. Daly, S. Bhaskar, R. Day, B. Anderson, A. K. Voss,
T. Thomas, L. G. Biesecker, P. Smith, A. Fryer, K. E. Chandler, B. Kerr, M. Tassabehji,
S. A. Lynch, M. Krajewska-Walasek, S. McKee, J. Smith, E. Sweeney, S. Mansour,
S. Mohammed, D. Donnai, and G. Black (2011). “Whole-exome-sequencing identifies
mutations in histone acetyltransferase gene KAT6B in individuals with the Say-
286
Barber-Biesecker variant of Ohdo syndrome”. The American Journal of Human Genetics
Vol. 89: 675–681.
Cristianini, N. and M. W. Hahn (2007). Introduction to Computational Genomics. Cam-
bridge University Press.
Czerkas, S. and A. Feduccia (2014). “Jurassic archosaur is a non-dinosaurian bird”.
Journal of Ornithology Vol. 155: 841–851.
De Jong, J., A. De Ronde, W. Keulen, M. Tersmette, and J. Goudsmit (1992). “Minimal Re-
quirements for the Human Immunodeficiency Virus Type 1 V3 Domain To Support
the Syncytium-Inducing Phenotype: Analysis by Single Amino Acid Substitution”.
Journal of Virology Vol. 66: 6777–6780.
DeRisi, J. L., V. R. Iyer, and P. O. Brown (1997). “Exploring the metabolic and genetic
control of gene expression on a genomic scale”. Science Vol. 278: 680–686.
Do, C. B. and S. Batzoglou (2008). “What is the expectation maximization algorithm?”
Nature Biotechnology Vol. 26: 897–899.
Eisen, M. B., P. T. Spellman, P. O. Brown, and D. Botstein (1998). “Cluster analysis and
display of genome-wide expression patterns”. Proceedings of the National Academy of
Sciences Vol. 95: 14863–14868.
Eng, J. K., A. L. McCormack, and J. R. Yates (1994). “An approach to correlate tandem
mass spectral data of peptides with amino acid sequences in a protein database”.
Journal of the American Society for Mass Spectrometry Vol. 5: 976 –989.
Felsenstein, J. (2004). Inferring Phylogenies. Sinauer Associates.
Ferragina, P. and G. Manzini (2000). “Opportunistic Data Structures with Applications”.
In: Proceedings of the 41st Annual Symposium on Foundations of Computer Science. IEEE
Computer Society, 390–398.
Gao, F., E. Bailes, D. L. Robertson, Y. Chen, C. M. Rodenburg, S. F. Michael, L. B.
Cummins, L. O. Arthur, M. Peeters, G. M. Shaw, P. M. Sharp, and B. H. Hahn (1999).
“Origin of HIV-1 in the chimpanzee Pan troglodytes troglodytes”. Nature Vol. 397:
436–441.
Green, R. E., E. L. Braun, J. Armstrong, D. Earl, N. Nguyen, G. Hickey, M. W. Vandewege,
J. A. St. John, S. Capella-Gutiérrez, T. A. Castoe, C. Kern, M. K. Fujita, J. C. Opazo,
J. Jurka, K. K. Kojima, J. Caballero, R. M. Hubley, A. F. Smit, R. N. Platt, C. A.
Lavoie, M. P. Ramakodi, J. W. Finger, A. Suh, S. R. Isberg, L. Miles, A. Y. Chong,
287
W. Jaratlerdsiri, J. Gongora, C. Moran, A. Iriarte, J. McCormack, S. C. Burgess, S.
V. Edwards, E. Lyons, C. Williams, M. Breen, J. T. Howard, C. R. Gresham, D. G.
Peterson, J. Schmitz, D. D. Pollock, D. Haussler, E. W. Triplett, G. Zhang, N. Irie, E. D.
Jarvis, C. A. Brochu, C. J. Schmidt, F. M. McCarthy, B. C. Faircloth, F. G. Hoffmann,
T. C. Glenn, T. Gabaldón, B. Paten, and D. A. Ray (2014). “Three crocodilian genomes
reveal ancestral patterns of evolution among archosaurs”. Science Vol. 346:
Hedges, S. and M. Schweitzer (1995). “Detecting dinosaur DNA”. Science Vol. 268: 1191–
1192.
Kellis, M., B. W. Birren, and E. S. Lander (2004). “Proof and evolutionary analysis of
ancient genome duplication in the yeast Saccharomyces cerevisiae”. Nature Vol. 428:
617–624.
Kim, S., M. J. Fitzgibbon, M. W. McIntosh, and P. A. Pevzner (2015). “Tyrannosaurus rex
protein sequencing controversy: contamination or treasure trove of ancient proteins?”
(unpublished manuscript).
Kim, S., N. Gupta, and P. A. Pevzner (2008). “Spectral Probabilities and Generating
Functions of Tandem Mass Spectra: A Strike against Decoy Databases”. Journal of
Proteome Research Vol. 7: PMID: 18597511, 3354–3363.
Kim, S., N. Gupta, N. Bandeira, and P. A. Pevzner (2009). “Spectral dictionaries: Inte-
grating de novo peptide sequencing with database search of tandem mass spectra.”
Molecular & Cellular Proteomics Vol. 8: 53–69.
Krogh, A., M. Brown, I. Saira Mian, K. Sjolander, and D. Haussler (1994). “Hidden
Markov Models in Computational Biology: Applications to Protein Modeling”.
Journal of Molecular Biology Vol. 235: 1501 –1531.
Lloyd, S. (1982). “Least squares quantization in PCM”. IEEE Transactions on Information
Theory Vol. 28: 129–137.
Manber, U. and G. Myers (1990). “Suffix Arrays: A New Method for On-line String
Searches”. In: Proceedings of the First Annual ACM-SIAM Symposium on Discrete
Algorithms. Society for Industrial and Applied Mathematics, 319–327.
Martin, N., E. Ruedi, R. LeDuc, F.-J. Sun, and G. Caetano-Anolles (2007). “Gene-
interleaving patterns of synteny in the Saccharomyces cerevisiae genome: Are they
proof of an ancient genome duplication event?” Biology Direct Vol. 2: 23.
288
Metzker, M. L., D. P. Mindell, X.-M. Liu, R. G. Ptak, R. A. Gibbs, and D. M. Hillis (2002).
“Molecular evidence of HIV-1 transmission in a criminal case”. Proceedings of the
National Academy of Sciences Vol. 99: 14292–14297.
Montgomery, S. B., D. Goode, E. Kvikstad, C. A. Albers, Z. Zhang, X. J. Mu, G. Ananda,
B. Howie, K. J. Karczewski, K. S. Smith, V. Anaya, R. Richardson, J. Davis, D. G.
MacArthur, A. Sidow, L. Duret, M. Gerstein, K. Markova, J. Marchini, G. A. McVean,
and G. Lunter (2013). “The origin, evolution and functional impact of short insertion-
deletion variants identified in 179 human genomes”. Genome Research Vol. 23: 749–
761.
O’Brien, S. J., W. G. Nash, D. E. Wildt, M. E. Bush, and R. E. Benveniste (1985). “A
molecular solution to the riddle of the giant panda’s phylogeny”. Nature Vol. 317:
140–144.
Ohno, S. (1970). Evolution by Gene Duplication. Springer-Verlag.
Pevzner, P. A., V. Dančík, and C. L. Tang (2000). “Mutation-Tolerant Protein Identification
by Mass Spectrometry”. Journal of Computational Biology Vol. 7: 777–787.
Pevzner, P. A., S. Kim, and J. Ng (2008). “Comment on ’Protein sequences from mastodon
and Tyrannosaurus rex revealed by mass spectrometry’”. Science Vol. 321: 1040.
Robinson, D. F. (1971). “Comparison of labeled trees with valency three”. Journal of
Combinatorial Theory, Series B Vol. 11: 105–119.
Saitou, N. and M. Nei (1987). “The neighbor-joining method: a new method for recon-
structing phylogenetic trees.” Mol Biol Evol Vol. 4: 406–425.
Sankoff, D. (1975). “Minimal Mutation Trees of Sequences”. SIAM Journal on Applied
Mathematics Vol. 28: 35–42.
Sokal, R. R. and C. D. Michener (1958). “A statistical method for evaluating systematic
relationships”. University of Kansas Scientific Bulletin Vol. 28: 1409–1438.
Studier, J. A. and K. J. Keppler (1988). “A note on the neighbor-joining algorithm of
Saitou and Nei.” Mol Biol Evol Vol. 5: 729–731.
Thomson, J. M., E. A. Gaucher, M. F. Burgan, D. W. De Kee, T. Li, J. P. Aris, and S. A.
Benner (2005). “Resurrecting ancestral alcohol dehydrogenases from yeast.” Nature
Genetics Vol. 37: 630–635.
289
Tuzun, E., A. J. Sharp, J. A. Bailey, R. Kaul, V. A. Morrison, L. M. Pertz, E. Haugen,
H. Hayden, D. Albertson, D. Pinkel, M. V. Olson, and E. E. Eichler (2005). “Fine-scale
structural variation of the human genome”. Nature Genetics Vol. 37: 727–732.
Viterbi, A. (1967). “Error bounds for convolutional codes and an asymptotically op-
timum decoding algorithm”. IEEE Transactions on Information Theory Vol. 13: 260–
269.
Weiner, P. (1973). “Linear Pattern Matching Algorithms”. In: Proceedings of the 14th
Annual Symposium on Switching and Automata Theory. IEEE Computer Society, 1–11.
Whiting, M. F., S. Bradler, and T. Maxwell (2003). “Loss and recovery of wings in stick
insects”. Nature Vol. 421: 264–267.
Wolfe, K. H. and D. C. Shields (1997). “Molecular evidence for an ancient duplication of
the entire yeast genome”. Nature Vol. 387: 708–713.
Woodward, N. Weyand, and M Bunnell (1994). “DNA sequence from Cretaceous period
bone fragments”. Science Vol. 266: 1229–1232.
Zaretskii, Z. A. (1965). “Constructing a tree on the basis of a set of distances between
the hanging vertices”. Uspekhi Mat. Nauk Vol. 20: 90–92.
Zuckerkandl, E. and L. Pauling (1965). “Molecules as documents of evolutionary history.”
Journal of Theoretical Biology Vol. 8: 357–366.
290
Image Courtesies
Figure 7.1 (right): Luc Viatour

Figure 7.14: Wikimedia Commons user Praveenp
Figure 7.15: Julie Langford (squirrel monkey), André Karwath (old world monkey),
Frans de Waal (chimp), Kabir Bakie (gorilla)
Figure 7.21: Wikimedia Commons user Drägüs (winged stick insect), L. Shyamal (wing-
less stick insect), Katka Nemčoková (giant centipede)
Figure 7.30: Frans de Waal (chimp), Kabir Bakie (gorilla)
Figure 7.36: Wikimedia Commons user Praveenp
Figure 7.41: Alan D. Wilson (polar bear), Mark Dumont (spectacled bear), Peter Meenen
(red panda), Wikimedia commons user Darkone (raccoon).
Figure 11.24: Matt Martyniuk
Figure 8.27: Guillaume Paumier (microarray)
Figure 10.27: Thomas Splettstoesser
Figure 10.28 (top): Wikimedia Commons user Dna 621
291

Phillip Compeau, Pavel Pevzner - Bioinformatics Algorithms - An Active Learning Approach. Vol2 (2015)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Phillip Compeau, Pavel Pevzner - Bioinformatics Algorithms - An Active Learning Approach. Vol2 (2015)

Uploaded by

Copyright:

Available Formats

Welcome!

2. CHARGING STATIONS provide additional insights on implementing the algo-

3. EXERCISE BREAKS offer “just in time” assessments testing your understanding

Phillip Compeau & Pavel Pevzner

Printed in the United States of America

First Printing, 2015

Library of Congress Control Number: 2015945208

Active Learning Publishers, LLC

List of Code Challenges xvii

About the Textbook xix

7 Which Animal Gave Us SARS? 2

8 How Did Yeast Become a Wine Maker? 68

9 How Do We Locate Disease-Causing Mutations? 120

10 Why Have Biologists Still Not Developed an HIV Vaccine? 178

11 Was T. rex Just a Big Chicken? 234

Image Courtesies 291

Meet the Authors

P HILLIP C OMPEAU is an Assistant Teaching Professor in the Compu-

PAVEL P EVZNER is Ronald R. Taylor Professor of Computer Science

V U N GO is a Ph. D student in Bioinformatics and Systems Biology at

M AX S HEN is a student in the Computational and Systems Biology

J EFFREY Y UAN is a graduate student in the Bioinformatics and Sys-

The Fastest Outbreak

Trouble at the Metropole Hotel

The evolution of SARS

did they use to construct this phylogeny?

Transforming Distance Matrices into Evolutionary Trees

Constructing a distance matrix from coronavirus genomes

to SARS-CoV. However, constructing a multiple alignment of entire viral genomes is

SPECIES ALIGNMENT DISTANCE MATRIX

Evolutionary trees as graphs

EXERCISE BREAK: Prove the following statements:

• Every tree with n nodes has n 1 edges.

owering owering worms

lophophorates rotifers roundworms

cartilaginous segmented mollusks chelicerates

Distance-based phylogeny construction

Distances Between Leaves Problem:

Input: A weighted tree with n leaves. 7A

Distance-Based Phylogeny Problem:

Input: A distance matrix.

Toward An Algorithm for Distance-Based Phylogeny Construction

A quest for neighboring leaves

(di,m + dk,m ) + (d j,m + dk,m ) (di,m + d j,m ) di,k + d j,k di,j

i di, k = di, m + dk, m k

di, j = di, m + dj, m m

j dj, k = dj, m + dk, m

di, j = di, m + dj, m m

• solve the Distance-Based Phylogeny problem for the smaller tree;

• add the previously removed leaves i and j back to the tree.

Computing limb lengths

Limb Length Problem:

Input: An additive distance matrix D and an integer j.

di,j = di, PARENT( j) + L IMB L ENGTH ( j)

Adding these two equations yields

di,j + d j,k = di, PARENT( j) + dk, PARENT( j) + 2 · L IMB L ENGTH ( j).

Because di, PARENT( j) + dk, PARENT( j) is equal to di,k , it follows that

di, PARENT( j) + dl, PARENT( j) di,l .

Combining this with the equation

di,j + d j, l = di,PARENT( j) + dl,PARENT( j) + 2 · L IMB L ENGTH ( j)

di,j + d j,l (di, PARENT( j) + dl, PARENT( j) )

EXERCISE BREAK: The proposed algorithm computes L IMB L ENGTH ( j) in

Trimming the tree